Merge pull request #1 from smthemex/PR

add new node

FoleyCrafter_node.py

@@ -0,0 +1,351 @@
+# !/usr/bin/env python
+# -*- coding: UTF-8 -*-
+
+import os
+import random
+import sys
+import cv2
+import numpy as np
+import torchaudio
+from PIL import Image
+
+import platform
+import subprocess
+import os.path as osp
+import soundfile as sf
+import torch
+import torchvision
+from huggingface_hub import snapshot_download, hf_hub_download
+from moviepy.editor import AudioFileClip, VideoFileClip
+
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+from .foleycrafter.models.onset import torch_utils
+from .foleycrafter.models.time_detector.model import VideoOnsetNet
+from .foleycrafter.pipelines.auffusion_pipeline import Generator, denormalize_spectrogram
+from. foleycrafter.utils.util import build_foleycrafter, read_frames_with_moviepy
+import folder_paths
+from comfy.utils import common_upscale
+
+vision_transform_list = [
+    torchvision.transforms.Resize((128, 128)),
+    torchvision.transforms.CenterCrop((112, 112)),
+    torchvision.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+]
+video_transform = torchvision.transforms.Compose(vision_transform_list)
+
+MAX_SEED = np.iinfo(np.int32).max
+
+ffmpeg_path = os.getenv('FFMPEG_PATH')
+if ffmpeg_path is None and platform.system() in ['Linux', 'Darwin']:
+    try:
+        result = subprocess.run(['which', 'ffmpeg'], capture_output=True, text=True)
+        if result.returncode == 0:
+            ffmpeg_path = result.stdout.strip()
+            print(f"FFmpeg is installed at: {ffmpeg_path}")
+        else:
+            print("FFmpeg is not installed. Please download ffmpeg-static and export to FFMPEG_PATH.")
+            print("For example: export FFMPEG_PATH=/you_path/ffmpeg-4.4-amd64-static")
+    except Exception as e:
+        pass
+
+if ffmpeg_path is not None and ffmpeg_path not in os.getenv('PATH'):
+    print("Adding FFMPEG_PATH to PATH")
+    os.environ["PATH"] = f"{ffmpeg_path}:{os.environ['PATH']}"
+
+device = ("cuda"
+    if torch.cuda.is_available()
+    else "mps" if torch.backends.mps.is_available() else "cpu"
+)
+
+weigths_current_path = os.path.join(folder_paths.models_dir, "foleycrafter")
+if not os.path.exists(weigths_current_path):
+    os.makedirs(weigths_current_path)
+
+weigths_semantic_current_path = os.path.join(weigths_current_path, "semantic")
+if not os.path.exists(weigths_semantic_current_path):
+    os.makedirs(weigths_semantic_current_path)
+
+weigths_vocoder_current_path = os.path.join(weigths_current_path, "vocoder")
+if not os.path.exists(weigths_vocoder_current_path):
+    os.makedirs(weigths_vocoder_current_path)
+
+
+def tensor_to_pil(tensor):
+    image_np = tensor.squeeze().mul(255).clamp(0, 255).byte().numpy()
+    image = Image.fromarray(image_np, mode='RGB')
+    return image
+
+
+def pil2narry(img):
+    narry = torch.from_numpy(np.array(img).astype(np.float32) / 255.0).unsqueeze(0)
+    return narry
+
+
+def narry_list(list_in):
+    for i in range(len(list_in)):
+        value = list_in[i]
+        modified_value = pil2narry(value)
+        list_in[i] = modified_value
+    return list_in
+
+
+def nomarl_upscale(img_tensor, width, height):
+    samples = img_tensor.movedim(-1, 1)
+    img = common_upscale(samples, width, height, "nearest-exact", "center")
+    samples = img.movedim(1, -1)
+    img_pil = tensor_to_pil(samples)
+    return img_pil
+
+
+def get_instance_path(path):
+    instance_path = os.path.normpath(path)
+    if sys.platform == 'win32':
+        instance_path = instance_path.replace('\\', "/")
+    return instance_path
+
+
+def download_weights(file_dir,repo_id,subfolder="",pt_name=""):
+    if subfolder:
+        file_path = os.path.join(file_dir,subfolder, pt_name)
+        sub_dir=os.path.join(file_dir,subfolder)
+        if not os.path.exists(sub_dir):
+            os.makedirs(sub_dir)
+        if not os.path.exists(file_path):
+            pt_path = hf_hub_download(
+                repo_id=repo_id,
+                subfolder=subfolder,
+                filename=pt_name,
+                local_dir = file_dir,
+            )
+        else:
+            pt_path=get_instance_path(file_path)
+        return pt_path
+    else:
+        file_path = os.path.join(file_dir, pt_name)
+        if not os.path.exists(file_dir):
+            os.makedirs(file_dir)
+        if not os.path.exists(file_path):
+            pt_path = hf_hub_download(
+                repo_id=repo_id,
+                filename=pt_name,
+                local_dir=file_dir,
+            )
+        else:
+            pt_path=get_instance_path(file_path)
+        return pt_path
+
+
+def build_models(repo,device,semantic_scale):
+    # download ckpt
+    if not repo:
+        repo = snapshot_download("auffusion/auffusion-full-no-adapter")
+    # pre models
+    download_weights(weigths_current_path, "ymzhang319/FoleyCrafter", subfolder="semantic",
+                         pt_name="semantic_adapter.bin")
+    download_weights(weigths_current_path, "ymzhang319/FoleyCrafter", subfolder="vocoder",
+                     pt_name="vocoder.pt")
+    download_weights(weigths_current_path, "ymzhang319/FoleyCrafter", subfolder="vocoder",
+                     pt_name="config.json")
+
+    # ckpt path
+    temporal_ckpt_path=download_weights(weigths_current_path, "ymzhang319/FoleyCrafter",
+                     pt_name="temporal_adapter.ckpt")
+    # load vocoder
+    vocoder = Generator.from_pretrained(weigths_vocoder_current_path).to(device)
+
+    # load time_detector
+    time_detector_ckpt=download_weights(weigths_current_path, "ymzhang319/FoleyCrafter",
+                     pt_name="timestamp_detector.pth.tar")
+
+    time_detector = VideoOnsetNet(False)
+    time_detector, _ = torch_utils.load_model(time_detector_ckpt, time_detector, device=device, strict=True)
+
+    # load adapters
+    pipe = build_foleycrafter(repo).to(device)
+    ckpt = torch.load(temporal_ckpt_path)
+
+    # load temporal adapter
+    if "state_dict" in ckpt.keys():
+        ckpt = ckpt["state_dict"]
+    load_gligen_ckpt = {}
+    for key, value in ckpt.items():
+        if key.startswith("module."):
+            load_gligen_ckpt[key[len("module.") :]] = value
+        else:
+            load_gligen_ckpt[key] = value
+    m, u = pipe.controlnet.load_state_dict(load_gligen_ckpt, strict=False)
+    print(f"### Control Net missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
+
+    # load semantic adapter
+    pipe.load_ip_adapter(weigths_semantic_current_path, subfolder="", weight_name="semantic_adapter.bin", image_encoder_folder=None)
+    pipe.set_ip_adapter_scale(semantic_scale)
+
+    return pipe, vocoder, time_detector
+
+
+class FoleyCrafter_LoadModel:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "repo_id":("STRING", {"default": "auffusion/auffusion-full-no-adapter"}),
+                "ip_repo":("STRING", {"default": "h94/IP-Adapter"}),
+                "semantic_scale":("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.1, "round": 0.01}),
+            }
+        }
+
+    RETURN_TYPES = ("MODEL","MODEL","MODEL","MODEL",)
+    RETURN_NAMES = ("pipe","vocoder","time_detector","image_encoder",)
+    FUNCTION = "main_loader"
+    CATEGORY = "FoleyCrafter"
+
+    def main_loader(self,repo_id,ip_repo,semantic_scale):
+        pipe, vocoder, time_detector=build_models(repo_id, device, semantic_scale)
+        image_encoder = CLIPVisionModelWithProjection.from_pretrained(ip_repo, subfolder="models/image_encoder").to(device)
+        return (pipe,vocoder, time_detector,image_encoder,)
+
+
+class FoleyCrafter_Sampler:
+    @classmethod
+    def INPUT_TYPES(s):
+        input_path = folder_paths.get_input_directory()
+        video_files = [f for f in os.listdir(input_path) if
+                       os.path.isfile(os.path.join(input_path, f)) and f.split('.')[-1] in ['webm', 'mp4', 'mkv','gif']]
+        return {
+            "required": {
+                "pipe": ("MODEL",),
+                "vocoder": ("MODEL",),
+                "time_detector": ("MODEL",),
+                "image_encoder": ("MODEL",),
+                "video_files": (["none"] + video_files,),
+                "prompt": ("STRING", {"multiline": True,"default": "1 girl"}),
+                "negative_prompt": ("STRING", {"multiline": True,"default": "bad quality"}),
+                "max_frame": ("INT", {"default": 150, "min": 64, "step": 1,"max": 500}),
+                "seeds": ("INT", {"default": 0, "min": 0, "max": 10000}),
+                "controlnet_scale": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 1.0, "step": 0.1, "round": 0.01}),
+                "steps": ("INT", {"default": 25, "min": 1, "max": 100}),
+                "sample_width": ("INT", {"default": 1024, "min": 128, "max": 1024, "step": 64, "display": "number"}),
+                "sample_height": ("INT", {"default": 256, "min": 128, "max": 1024, "step": 64, "display": "number"}),
+                "video_dubbing": ("BOOLEAN", {"default": False},), }
+        }
+    
+    RETURN_TYPES = ("IMAGE", "AUDIO", "FLOAT")
+    RETURN_NAMES = ("image", "audio", "frame_rate",)
+    FUNCTION = "fc_main"
+    CATEGORY = "FoleyCrafter"
+    
+    def run_inference(self,pipe, vocoder, time_detector,controlnet_scale,seeds,image_encoder,video_files,max_frame,prompt,negative_prompt,steps,width,height,video_dubbing):
+
+        generator = torch.Generator(device=device)
+        generator.manual_seed(seeds)
+        image_processor = CLIPImageProcessor()
+
+        source_video = cv2.VideoCapture(video_files)
+        fps = source_video.get(cv2.CAP_PROP_FPS)
+        v_width = int(source_video.get(cv2.CAP_PROP_FRAME_WIDTH))  # 获取视频的fps w h
+        v_height = int(source_video.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        source_video.release()
+        
+        with torch.no_grad():
+            print(f" >>> Begin Inference: {video_files} <<< ")
+            frames, duration,frames_list = read_frames_with_moviepy(video_files, max_frame_nums=max_frame)
+            
+            time_frames = torch.FloatTensor(frames).permute(0, 3, 1, 2)
+            time_frames = video_transform(time_frames)
+            time_frames = {"frames": time_frames.unsqueeze(0).permute(0, 2, 1, 3, 4)}
+            preds = time_detector(time_frames)
+            preds = torch.sigmoid(preds)
+            
+            # duration
+            # import ipdb; ipdb.set_trace()
+            time_condition = [
+                -1 if preds[0][int(i / (1024 / 10 * duration) * 150)] < 0.5 else 1
+                for i in range(int(1024 / 10 * duration))
+            ]
+            time_condition = time_condition + [-1] * (1024 - len(time_condition))
+            # w -> b c h w
+            time_condition = (
+                torch.FloatTensor(time_condition)
+                .unsqueeze(0)
+                .unsqueeze(0)
+                .unsqueeze(0)
+                .repeat(1, 1, 256, 1)
+                .to("cuda")
+            )
+            images = image_processor(images=frames, return_tensors="pt").to(device)
+            image_embeddings = image_encoder(**images).image_embeds
+            image_embeddings = torch.mean(image_embeddings, dim=0, keepdim=True).unsqueeze(0).unsqueeze(0)
+            neg_image_embeddings = torch.zeros_like(image_embeddings)
+            image_embeddings = torch.cat([neg_image_embeddings, image_embeddings], dim=1)
+            
+            sample = pipe(
+                prompt=prompt,
+                negative_prompt=negative_prompt,
+                ip_adapter_image_embeds=image_embeddings,
+                image=time_condition,
+                # audio_length_in_s=10,
+                controlnet_conditioning_scale=controlnet_scale,
+                num_inference_steps=steps,
+                height=height,
+                width=width,
+                output_type="pt",
+                generator=generator,
+                # guidance_scale=0,
+            )
+            
+            audio_img = sample.images[0]
+            audio = denormalize_spectrogram(audio_img)
+            audio = vocoder.inference(audio, lengths=160000)[0]
+            
+            waveform = audio[: int(duration * 16000)]
+            # save
+            rand_file_prefix = ''.join(random.choice("0123456789") for _ in range(6))
+            save_path = osp.join(folder_paths.output_directory, f"{rand_file_prefix}.wav")
+            sf.write(save_path, waveform, 16000)
+            
+            # output
+            waveform, sample_rate = torchaudio.load(save_path)
+            audio_output = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate}
+            
+            if video_dubbing:
+                output_video_path = os.path.join(folder_paths.output_directory, f"{rand_file_prefix}_audio.mp4")
+                video_clip = VideoFileClip(video_files)
+                audio_clip = AudioFileClip(save_path)
+                final_clip = video_clip.set_audio(audio_clip)
+                final_clip.write_videofile(
+                    output_video_path,
+                    codec="libx264", audio_codec="aac")
+                print(f"saving {output_video_path}")
+                video_clip.reader.close()
+                audio_clip.close()
+                final_clip.reader.close()
+                
+            return frames_list, audio_output, fps,v_width,v_height
+    
+            
+    def fc_main(self, pipe, vocoder, time_detector,image_encoder,video_files, prompt,negative_prompt, seeds,max_frame,controlnet_scale, steps,sample_width,sample_height, video_dubbing, ):
+        if video_files != "none":
+            video_files=osp.join(folder_paths.input_directory, video_files)
+            images,audio,fps,v_width,v_height=self.run_inference( pipe, vocoder, time_detector,controlnet_scale,seeds,image_encoder,video_files,max_frame,prompt,negative_prompt,steps,sample_width,sample_height,video_dubbing)
+        else:
+            raise "need video file"
+        frame_rate=float(fps)
+        gen = narry_list(images)
+        images = torch.from_numpy(np.fromiter(gen, np.dtype((np.float32, (v_height, v_width, 3)))))
+        torch.cuda.empty_cache()
+        
+        return (images,audio,frame_rate,)
+
+
+
+NODE_CLASS_MAPPINGS = {
+    "FoleyCrafter_LoadModel":FoleyCrafter_LoadModel,
+    "FoleyCrafter_Sampler": FoleyCrafter_Sampler
+}
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "FoleyCrafter_LoadModel":"FoleyCrafter_LoadModel",
+    "FoleyCrafter_Sampler": "FoleyCrafter_Sampler",
+}

__init__.py

@@ -0,0 +1,4 @@
+
+from .FoleyCrafter_node import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]

foleycrafter/data/__init__.py

@@ -0,0 +1,27 @@
+from .dataset import AudioSetStrong, CPU_Unpickler, VGGSound, dynamic_range_compression, get_mel, zero_rank_print
+from .video_transforms import (
+    CenterCropVideo,
+    KineticsRandomCropResizeVideo,
+    NormalizeVideo,
+    RandomHorizontalFlipVideo,
+    TemporalRandomCrop,
+    ToTensorVideo,
+    UCFCenterCropVideo,
+)
+
+
+__all__ = [
+    "zero_rank_print",
+    "get_mel",
+    "dynamic_range_compression",
+    "CPU_Unpickler",
+    "AudioSetStrong",
+    "VGGSound",
+    "UCFCenterCropVideo",
+    "KineticsRandomCropResizeVideo",
+    "CenterCropVideo",
+    "NormalizeVideo",
+    "ToTensorVideo",
+    "RandomHorizontalFlipVideo",
+    "TemporalRandomCrop",
+]

foleycrafter/data/dataset.py

@@ -0,0 +1,155 @@
+import glob
+import io
+import pickle
+import random
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torchaudio
+import torchvision.transforms as transforms
+from torch.utils.data.dataset import Dataset
+
+
+def zero_rank_print(s):
+    if (not dist.is_initialized()) or (dist.is_initialized() and dist.get_rank() == 0):
+        print("### " + s, flush=True)
+
+
+@torch.no_grad()
+def get_mel(audio_data, audio_cfg):
+    # mel shape: (n_mels, T)
+    mel = torchaudio.transforms.MelSpectrogram(
+        sample_rate=audio_cfg["sample_rate"],
+        n_fft=audio_cfg["window_size"],
+        win_length=audio_cfg["window_size"],
+        hop_length=audio_cfg["hop_size"],
+        center=True,
+        pad_mode="reflect",
+        power=2.0,
+        norm=None,
+        onesided=True,
+        n_mels=64,
+        f_min=audio_cfg["fmin"],
+        f_max=audio_cfg["fmax"],
+    ).to(audio_data.device)
+    mel = mel(audio_data)
+    # we use log mel spectrogram as input
+    mel = torchaudio.transforms.AmplitudeToDB(top_db=None)(mel)
+    return mel  # (T, n_mels)
+
+
+def dynamic_range_compression(x, normalize_fun=torch.log, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return normalize_fun(torch.clamp(x, min=clip_val) * C)
+
+
+class CPU_Unpickler(pickle.Unpickler):
+    def find_class(self, module, name):
+        if module == "torch.storage" and name == "_load_from_bytes":
+            return lambda b: torch.load(io.BytesIO(b), map_location="cpu")
+        else:
+            return super().find_class(module, name)
+
+
+class AudioSetStrong(Dataset):
+    # read feature and audio
+    def __init__(
+        self,
+        data_path="data/AudioSetStrong/train/feature",
+        video_path="data/AudioSetStrong/train/video",
+    ):
+        super().__init__()
+        self.data_path = data_path
+        self.data_list = list(self.data_path)
+        self.length = len(self.data_list)
+        # get video feature
+        self.video_path = video_path
+        vision_transform_list = [
+            transforms.Resize((128, 128)),
+            transforms.CenterCrop((112, 112)),
+            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+        ]
+        self.video_transform = transforms.Compose(vision_transform_list)
+
+    def get_batch(self, idx):
+        embeds = self.data_list[idx]
+        mel = embeds["mel"]
+        save_bsz = mel.shape[0]
+        audio_info = embeds["audio_info"]
+        text_embeds = embeds["text_embeds"]
+
+        # audio_info['label_list'] = np.array(audio_info['label_list'])
+        audio_info_array = np.array(audio_info["label_list"])
+        prompts = []
+        for i in range(save_bsz):
+            prompts.append(", ".join(audio_info_array[i, : audio_info["event_num"][i]].tolist()))
+
+        return mel, audio_info, text_embeds, prompts
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        while True:
+            try:
+                mel, audio_info, text_embeds, prompts, videos = self.get_batch(idx)
+                break
+            except Exception:
+                zero_rank_print(" >>> load error <<<")
+                idx = random.randint(0, self.length - 1)
+        sample = {
+            "mel": mel,
+            "audio_info": audio_info,
+            "text_embeds": text_embeds,
+            "prompts": prompts,
+            "videos": videos,
+        }
+        return sample
+
+
+class VGGSound(Dataset):
+    # read feature and audio
+    def __init__(
+        self,
+        data_path="data/VGGSound/train/video",
+        visual_data_path="data/VGGSound/train/feature",
+    ):
+        super().__init__()
+        self.data_path = data_path
+        self.visual_data_path = visual_data_path
+        self.embeds_list = glob.glob(f"{self.data_path}/*.pt")
+        self.visual_list = glob.glob(f"{self.visual_data_path}/*.pt")
+        self.length = len(self.embeds_list)
+
+    def get_batch(self, idx):
+        embeds = torch.load(self.embeds_list[idx], map_location="cpu")
+        visual_embeds = torch.load(self.visual_list[idx], map_location="cpu")
+
+        # audio_embeds  = embeds['audio_embeds']
+        visual_embeds = visual_embeds["visual_embeds"]
+        # video_name = embeds["video_name"]
+        text = embeds["text"]
+        mel = embeds["mel"]
+
+        audio = mel
+
+        return visual_embeds, audio, text
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        while True:
+            try:
+                visual_embeds, audio, text = self.get_batch(idx)
+                break
+            except Exception:
+                zero_rank_print("load error")
+                idx = random.randint(0, self.length - 1)
+        sample = {"visual_embeds": visual_embeds, "audio": audio, "text": text}
+        return sample

foleycrafter/data/video_transforms.py

@@ -0,0 +1,404 @@
+import numbers
+import random
+
+import torch
+
+
+def _is_tensor_video_clip(clip):
+    if not torch.is_tensor(clip):
+        raise TypeError("clip should be Tensor. Got %s" % type(clip))
+
+    if not clip.ndimension() == 4:
+        raise ValueError("clip should be 4D. Got %dD" % clip.dim())
+
+    return True
+
+
+def crop(clip, i, j, h, w):
+    """
+    Args:
+        clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+    """
+    if len(clip.size()) != 4:
+        raise ValueError("clip should be a 4D tensor")
+    return clip[..., i : i + h, j : j + w]
+
+
+def resize(clip, target_size, interpolation_mode):
+    if len(target_size) != 2:
+        raise ValueError(f"target size should be tuple (height, width), instead got {target_size}")
+    return torch.nn.functional.interpolate(clip, size=target_size, mode=interpolation_mode, align_corners=False)
+
+
+def resize_scale(clip, target_size, interpolation_mode):
+    if len(target_size) != 2:
+        raise ValueError(f"target size should be tuple (height, width), instead got {target_size}")
+    _, _, H, W = clip.shape
+    scale_ = target_size[0] / min(H, W)
+    return torch.nn.functional.interpolate(clip, scale_factor=scale_, mode=interpolation_mode, align_corners=False)
+
+
+def resized_crop(clip, i, j, h, w, size, interpolation_mode="bilinear"):
+    """
+    Do spatial cropping and resizing to the video clip
+    Args:
+        clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        i (int): i in (i,j) i.e coordinates of the upper left corner.
+        j (int): j in (i,j) i.e coordinates of the upper left corner.
+        h (int): Height of the cropped region.
+        w (int): Width of the cropped region.
+        size (tuple(int, int)): height and width of resized clip
+    Returns:
+        clip (torch.tensor): Resized and cropped clip. Size is (T, C, H, W)
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    clip = crop(clip, i, j, h, w)
+    clip = resize(clip, size, interpolation_mode)
+    return clip
+
+
+def center_crop(clip, crop_size):
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    h, w = clip.size(-2), clip.size(-1)
+    th, tw = crop_size
+    if h < th or w < tw:
+        raise ValueError("height and width must be no smaller than crop_size")
+
+    i = int(round((h - th) / 2.0))
+    j = int(round((w - tw) / 2.0))
+    return crop(clip, i, j, th, tw)
+
+
+def random_shift_crop(clip):
+    """
+    Slide along the long edge, with the short edge as crop size
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    h, w = clip.size(-2), clip.size(-1)
+
+    if h <= w:
+        # long_edge = w
+        short_edge = h
+    else:
+        # long_edge = h
+        short_edge = w
+
+    th, tw = short_edge, short_edge
+
+    i = torch.randint(0, h - th + 1, size=(1,)).item()
+    j = torch.randint(0, w - tw + 1, size=(1,)).item()
+    return crop(clip, i, j, th, tw)
+
+
+def to_tensor(clip):
+    """
+    Convert tensor data type from uint8 to float, divide value by 255.0 and
+    permute the dimensions of clip tensor
+    Args:
+        clip (torch.tensor, dtype=torch.uint8): Size is (T, C, H, W)
+    Return:
+        clip (torch.tensor, dtype=torch.float): Size is (T, C, H, W)
+    """
+    _is_tensor_video_clip(clip)
+    if not clip.dtype == torch.uint8:
+        raise TypeError("clip tensor should have data type uint8. Got %s" % str(clip.dtype))
+    # return clip.float().permute(3, 0, 1, 2) / 255.0
+    return clip.float() / 255.0
+
+
+def normalize(clip, mean, std, inplace=False):
+    """
+    Args:
+        clip (torch.tensor): Video clip to be normalized. Size is (T, C, H, W)
+        mean (tuple): pixel RGB mean. Size is (3)
+        std (tuple): pixel standard deviation. Size is (3)
+    Returns:
+        normalized clip (torch.tensor): Size is (T, C, H, W)
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    if not inplace:
+        clip = clip.clone()
+    mean = torch.as_tensor(mean, dtype=clip.dtype, device=clip.device)
+    print(mean)
+    std = torch.as_tensor(std, dtype=clip.dtype, device=clip.device)
+    clip.sub_(mean[:, None, None, None]).div_(std[:, None, None, None])
+    return clip
+
+
+def hflip(clip):
+    """
+    Args:
+        clip (torch.tensor): Video clip to be normalized. Size is (T, C, H, W)
+    Returns:
+        flipped clip (torch.tensor): Size is (T, C, H, W)
+    """
+    if not _is_tensor_video_clip(clip):
+        raise ValueError("clip should be a 4D torch.tensor")
+    return clip.flip(-1)
+
+
+class RandomCropVideo:
+    def __init__(self, size):
+        if isinstance(size, numbers.Number):
+            self.size = (int(size), int(size))
+        else:
+            self.size = size
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            torch.tensor: randomly cropped video clip.
+                size is (T, C, OH, OW)
+        """
+        i, j, h, w = self.get_params(clip)
+        return crop(clip, i, j, h, w)
+
+    def get_params(self, clip):
+        h, w = clip.shape[-2:]
+        th, tw = self.size
+
+        if h < th or w < tw:
+            raise ValueError(f"Required crop size {(th, tw)} is larger than input image size {(h, w)}")
+
+        if w == tw and h == th:
+            return 0, 0, h, w
+
+        i = torch.randint(0, h - th + 1, size=(1,)).item()
+        j = torch.randint(0, w - tw + 1, size=(1,)).item()
+
+        return i, j, th, tw
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size})"
+
+
+class UCFCenterCropVideo:
+    def __init__(
+        self,
+        size,
+        interpolation_mode="bilinear",
+    ):
+        if isinstance(size, tuple):
+            if len(size) != 2:
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
+            self.size = size
+        else:
+            self.size = (size, size)
+
+        self.interpolation_mode = interpolation_mode
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            torch.tensor: scale resized / center cropped video clip.
+                size is (T, C, crop_size, crop_size)
+        """
+        clip_resize = resize_scale(clip=clip, target_size=self.size, interpolation_mode=self.interpolation_mode)
+        clip_center_crop = center_crop(clip_resize, self.size)
+        return clip_center_crop
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}"
+
+
+class KineticsRandomCropResizeVideo:
+    """
+    Slide along the long edge, with the short edge as crop size. And resie to the desired size.
+    """
+
+    def __init__(
+        self,
+        size,
+        interpolation_mode="bilinear",
+    ):
+        if isinstance(size, tuple):
+            if len(size) != 2:
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
+            self.size = size
+        else:
+            self.size = (size, size)
+
+        self.interpolation_mode = interpolation_mode
+
+    def __call__(self, clip):
+        clip_random_crop = random_shift_crop(clip)
+        clip_resize = resize(clip_random_crop, self.size, self.interpolation_mode)
+        return clip_resize
+
+
+class CenterCropVideo:
+    def __init__(
+        self,
+        size,
+        interpolation_mode="bilinear",
+    ):
+        if isinstance(size, tuple):
+            if len(size) != 2:
+                raise ValueError(f"size should be tuple (height, width), instead got {size}")
+            self.size = size
+        else:
+            self.size = (size, size)
+
+        self.interpolation_mode = interpolation_mode
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W)
+        Returns:
+            torch.tensor: center cropped video clip.
+                size is (T, C, crop_size, crop_size)
+        """
+        clip_center_crop = center_crop(clip, self.size)
+        return clip_center_crop
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}"
+
+
+class NormalizeVideo:
+    """
+    Normalize the video clip by mean subtraction and division by standard deviation
+    Args:
+        mean (3-tuple): pixel RGB mean
+        std (3-tuple): pixel RGB standard deviation
+        inplace (boolean): whether do in-place normalization
+    """
+
+    def __init__(self, mean, std, inplace=False):
+        self.mean = mean
+        self.std = std
+        self.inplace = inplace
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): video clip must be normalized. Size is (C, T, H, W)
+        """
+        return normalize(clip, self.mean, self.std, self.inplace)
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(mean={self.mean}, std={self.std}, inplace={self.inplace})"
+
+
+class ToTensorVideo:
+    """
+    Convert tensor data type from uint8 to float, divide value by 255.0 and
+    permute the dimensions of clip tensor
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor, dtype=torch.uint8): Size is (T, C, H, W)
+        Return:
+            clip (torch.tensor, dtype=torch.float): Size is (T, C, H, W)
+        """
+        return to_tensor(clip)
+
+    def __repr__(self) -> str:
+        return self.__class__.__name__
+
+
+class RandomHorizontalFlipVideo:
+    """
+    Flip the video clip along the horizontal direction with a given probability
+    Args:
+        p (float): probability of the clip being flipped. Default value is 0.5
+    """
+
+    def __init__(self, p=0.5):
+        self.p = p
+
+    def __call__(self, clip):
+        """
+        Args:
+            clip (torch.tensor): Size is (T, C, H, W)
+        Return:
+            clip (torch.tensor): Size is (T, C, H, W)
+        """
+        if random.random() < self.p:
+            clip = hflip(clip)
+        return clip
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(p={self.p})"
+
+
+#  ------------------------------------------------------------
+#  ---------------------  Sampling  ---------------------------
+#  ------------------------------------------------------------
+class TemporalRandomCrop(object):
+    """Temporally crop the given frame indices at a random location.
+
+    Args:
+            size (int): Desired length of frames will be seen in the model.
+    """
+
+    def __init__(self, size):
+        self.size = size
+
+    def __call__(self, total_frames):
+        rand_end = max(0, total_frames - self.size - 1)
+        begin_index = random.randint(0, rand_end)
+        end_index = min(begin_index + self.size, total_frames)
+        return begin_index, end_index
+
+
+if __name__ == "__main__":
+    import os
+
+    import numpy as np
+    import torchvision.io as io
+    from torchvision import transforms
+    from torchvision.utils import save_image
+
+    vframes, aframes, info = io.read_video(filename="./v_Archery_g01_c03.avi", pts_unit="sec", output_format="TCHW")
+
+    trans = transforms.Compose(
+        [
+            ToTensorVideo(),
+            RandomHorizontalFlipVideo(),
+            UCFCenterCropVideo(512),
+            # NormalizeVideo(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
+        ]
+    )
+
+    target_video_len = 32
+    frame_interval = 1
+    total_frames = len(vframes)
+    print(total_frames)
+
+    temporal_sample = TemporalRandomCrop(target_video_len * frame_interval)
+
+    # Sampling video frames
+    start_frame_ind, end_frame_ind = temporal_sample(total_frames)
+    # print(start_frame_ind)
+    # print(end_frame_ind)
+    assert end_frame_ind - start_frame_ind >= target_video_len
+    frame_indice = np.linspace(start_frame_ind, end_frame_ind - 1, target_video_len, dtype=int)
+
+    select_vframes = vframes[frame_indice]
+
+    select_vframes_trans = trans(select_vframes)
+
+    select_vframes_trans_int = ((select_vframes_trans * 0.5 + 0.5) * 255).to(dtype=torch.uint8)
+
+    io.write_video("./test.avi", select_vframes_trans_int.permute(0, 2, 3, 1), fps=8)
+
+    for i in range(target_video_len):
+        save_image(
+            select_vframes_trans[i], os.path.join("./test000", "%04d.png" % i), normalize=True, value_range=(-1, 1)
+        )

foleycrafter/models/adapters/attention_processor.py

@@ -0,0 +1,655 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.utils import logging
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class AttnProcessor(nn.Module):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = attn.head_to_batch_dim(ip_key)
+        ip_value = attn.head_to_batch_dim(ip_value)
+
+        ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+        self.attn_map = ip_attention_probs
+        ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0WithProjection(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.before_proj_size = 1024
+        self.after_proj_size = 768
+        self.visual_proj = nn.Linear(self.before_proj_size, self.after_proj_size)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+        # encoder_hidden_states = self.visual_proj(encoder_hidden_states)
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        ip_hidden_states = F.scaled_dot_product_attention(
+            query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+        with torch.no_grad():
+            self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
+            # print(self.attn_map.shape)
+
+        ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+## for controlnet
+class CNAttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(self, num_tokens=4):
+        self.num_tokens = num_tokens
+
+    def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class CNAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self, num_tokens=4):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.num_tokens = num_tokens
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

foleycrafter/models/adapters/ip_adapter.py

@@ -0,0 +1,189 @@
+import torch
+import torch.nn as nn
+
+
+class IPAdapter(torch.nn.Module):
+    """IP-Adapter"""
+
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj_model = image_proj_model
+        self.adapter_modules = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj_model(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj_model.load_state_dict(state_dict["image_proj"], strict=True)
+        self.adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=True)
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_modules did not change!"
+
+        print(f"Successfully loaded weights from checkpoint {ckpt_path}")
+
+
+class ImageProjModel(torch.nn.Module):
+    """Projection Model"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class MLPProjModel(torch.nn.Module):
+    """SD model with image prompt"""
+
+    def zero_initialize(module):
+        for param in module.parameters():
+            param.data.zero_()
+
+    def zero_initialize_last_layer(module):
+        last_layer = None
+        for module_name, layer in module.named_modules():
+            if isinstance(layer, torch.nn.Linear):
+                last_layer = layer
+
+        if last_layer is not None:
+            last_layer.weight.data.zero_()
+            last_layer.bias.data.zero_()
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
+            torch.nn.LayerNorm(cross_attention_dim),
+        )
+        # zero initialize the last layer
+        # self.zero_initialize_last_layer()
+
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class V2AMapperMLP(torch.nn.Module):
+    def __init__(self, cross_attention_dim=512, clip_embeddings_dim=512, mult=4):
+        super().__init__()
+        self.proj = torch.nn.Sequential(
+            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim * mult),
+            torch.nn.GELU(),
+            torch.nn.Linear(clip_embeddings_dim * mult, cross_attention_dim),
+            torch.nn.LayerNorm(cross_attention_dim),
+        )
+
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class TimeProjModel(torch.nn.Module):
+    def __init__(self, positive_len, out_dim, feature_type="text-only", frame_nums: int = 64):
+        super().__init__()
+        self.positive_len = positive_len
+        self.out_dim = out_dim
+
+        self.position_dim = frame_nums
+
+        if isinstance(out_dim, tuple):
+            out_dim = out_dim[0]
+
+        if feature_type == "text-only":
+            self.linears = nn.Sequential(
+                nn.Linear(self.positive_len + self.position_dim, 512),
+                nn.SiLU(),
+                nn.Linear(512, 512),
+                nn.SiLU(),
+                nn.Linear(512, out_dim),
+            )
+            self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+
+        elif feature_type == "text-image":
+            self.linears_text = nn.Sequential(
+                nn.Linear(self.positive_len + self.position_dim, 512),
+                nn.SiLU(),
+                nn.Linear(512, 512),
+                nn.SiLU(),
+                nn.Linear(512, out_dim),
+            )
+            self.linears_image = nn.Sequential(
+                nn.Linear(self.positive_len + self.position_dim, 512),
+                nn.SiLU(),
+                nn.Linear(512, 512),
+                nn.SiLU(),
+                nn.Linear(512, out_dim),
+            )
+            self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+            self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+
+        # self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim]))
+
+    def forward(
+        self,
+        boxes,
+        masks,
+        positive_embeddings=None,
+    ):
+        masks = masks.unsqueeze(-1)
+
+        # # embedding position (it may includes padding as placeholder)
+        # xyxy_embedding = self.fourier_embedder(boxes)  # B*N*4 -> B*N*C
+
+        # # learnable null embedding
+        # xyxy_null = self.null_position_feature.view(1, 1, -1)
+
+        # # replace padding with learnable null embedding
+        # xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
+
+        time_embeds = boxes
+
+        # positionet with text only information
+        if positive_embeddings is not None:
+            # learnable null embedding
+            positive_null = self.null_positive_feature.view(1, 1, -1)
+
+            # replace padding with learnable null embedding
+            positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null
+
+            objs = self.linears(torch.cat([positive_embeddings, time_embeds], dim=-1))
+
+        # positionet with text and image information
+        else:
+            raise NotImplementedError
+
+        return objs

foleycrafter/models/adapters/resampler.py

@@ -0,0 +1,158 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
+
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        max_seq_len: int = 257,  # CLIP tokens + CLS token
+        apply_pos_emb: bool = False,
+        num_latents_mean_pooled: int = 0,  # number of latents derived from mean pooled representation of the sequence
+    ):
+        super().__init__()
+        self.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool))
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+def masked_mean(t, *, dim, mask=None):
+    if mask is None:
+        return t.mean(dim=dim)
+
+    denom = mask.sum(dim=dim, keepdim=True)
+    mask = rearrange(mask, "b n -> b n 1")
+    masked_t = t.masked_fill(~mask, 0.0)
+
+    return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)

foleycrafter/models/adapters/transformer.py

@@ -0,0 +1,352 @@
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+
+class Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, hidden_size, num_attention_heads, attention_head_dim, attention_dropout=0.0):
+        super().__init__()
+        self.embed_dim = hidden_size
+        self.num_heads = num_attention_heads
+        self.head_dim = attention_head_dim
+
+        self.scale = self.head_dim**-0.5
+        self.dropout = attention_dropout
+
+        self.inner_dim = self.head_dim * self.num_heads
+
+        self.k_proj = nn.Linear(self.embed_dim, self.inner_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.inner_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.inner_dim)
+        self.out_proj = nn.Linear(self.inner_dim, self.embed_dim)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scale
+        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        # apply the causal_attention_mask first
+        if causal_attention_mask is not None:
+            if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {causal_attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, tgt_len, self.inner_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped
+
+
+class MLP(nn.Module):
+    def __init__(self, hidden_size, intermediate_size, mult=4):
+        super().__init__()
+        self.activation_fn = nn.SiLU()
+        self.fc1 = nn.Linear(hidden_size, intermediate_size * mult)
+        self.fc2 = nn.Linear(intermediate_size * mult, hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class Transformer(nn.Module):
+    def __init__(self, depth=12):
+        super().__init__()
+        self.layers = nn.ModuleList([TransformerBlock() for _ in range(depth)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor = None,
+        causal_attention_mask: torch.Tensor = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        for layer in self.layers:
+            hidden_states = layer(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                causal_attention_mask=causal_attention_mask,
+                output_attentions=output_attentions,
+            )
+
+        return hidden_states
+
+
+class TransformerBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size=512,
+        num_attention_heads=12,
+        attention_head_dim=64,
+        attention_dropout=0.0,
+        dropout=0.0,
+        eps=1e-5,
+    ):
+        super().__init__()
+        self.embed_dim = hidden_size
+        self.self_attn = Attention(
+            hidden_size=hidden_size, num_attention_heads=num_attention_heads, attention_head_dim=attention_head_dim
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=eps)
+        self.mlp = MLP(hidden_size=hidden_size, intermediate_size=hidden_size)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor = None,
+        causal_attention_mask: torch.Tensor = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs[0]
+
+
+class DiffusionTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size=512,
+        num_attention_heads=12,
+        attention_head_dim=64,
+        attention_dropout=0.0,
+        dropout=0.0,
+        eps=1e-5,
+    ):
+        super().__init__()
+        self.embed_dim = hidden_size
+        self.self_attn = Attention(
+            hidden_size=hidden_size, num_attention_heads=num_attention_heads, attention_head_dim=attention_head_dim
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=eps)
+        self.mlp = MLP(hidden_size=hidden_size, intermediate_size=hidden_size)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=eps)
+        self.output_token = nn.Parameter(torch.randn(1, hidden_size))
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor = None,
+        causal_attention_mask: torch.Tensor = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        output_token = self.output_token.unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)
+        hidden_states = torch.cat([output_token, hidden_states], dim=1)
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs[0][:, 0:1, ...]
+
+
+class V2AMapperMLP(nn.Module):
+    def __init__(self, input_dim=512, output_dim=512, expansion_rate=4):
+        super().__init__()
+        self.linear = nn.Linear(input_dim, input_dim * expansion_rate)
+        self.silu = nn.SiLU()
+        self.layer_norm = nn.LayerNorm(input_dim * expansion_rate)
+        self.linear2 = nn.Linear(input_dim * expansion_rate, output_dim)
+
+    def forward(self, x):
+        x = self.linear(x)
+        x = self.silu(x)
+        x = self.layer_norm(x)
+        x = self.linear2(x)
+
+        return x
+
+
+class ImageProjModel(torch.nn.Module):
+    """Projection Model"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+        self.zero_initialize_last_layer()
+
+    def zero_initialize_last_layer(module):
+        last_layer = None
+        for module_name, layer in module.named_modules():
+            if isinstance(layer, torch.nn.Linear):
+                last_layer = layer
+
+        if last_layer is not None:
+            last_layer.weight.data.zero_()
+            last_layer.bias.data.zero_()
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class VisionAudioAdapter(torch.nn.Module):
+    def __init__(
+        self,
+        embedding_size=768,
+        expand_dim=4,
+        token_num=4,
+    ):
+        super().__init__()
+
+        self.mapper = V2AMapperMLP(
+            embedding_size,
+            embedding_size,
+            expansion_rate=expand_dim,
+        )
+
+        self.proj = ImageProjModel(
+            cross_attention_dim=embedding_size,
+            clip_embeddings_dim=embedding_size,
+            clip_extra_context_tokens=token_num,
+        )
+
+    def forward(self, image_embeds):
+        image_embeds = self.mapper(image_embeds)
+        image_embeds = self.proj(image_embeds)
+        return image_embeds

foleycrafter/models/adapters/utils.py

@@ -0,0 +1,86 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+from PIL import Image
+
+
+attn_maps = {}
+
+
+def hook_fn(name):
+    def forward_hook(module, input, output):
+        if hasattr(module.processor, "attn_map"):
+            attn_maps[name] = module.processor.attn_map
+            del module.processor.attn_map
+
+    return forward_hook
+
+
+def register_cross_attention_hook(unet):
+    for name, module in unet.named_modules():
+        if name.split(".")[-1].startswith("attn2"):
+            module.register_forward_hook(hook_fn(name))
+
+    return unet
+
+
+def upscale(attn_map, target_size):
+    attn_map = torch.mean(attn_map, dim=0)
+    attn_map = attn_map.permute(1, 0)
+    temp_size = None
+
+    for i in range(0, 5):
+        scale = 2**i
+        if (target_size[0] // scale) * (target_size[1] // scale) == attn_map.shape[1] * 64:
+            temp_size = (target_size[0] // (scale * 8), target_size[1] // (scale * 8))
+            break
+
+    assert temp_size is not None, "temp_size cannot is None"
+
+    attn_map = attn_map.view(attn_map.shape[0], *temp_size)
+
+    attn_map = F.interpolate(
+        attn_map.unsqueeze(0).to(dtype=torch.float32), size=target_size, mode="bilinear", align_corners=False
+    )[0]
+
+    attn_map = torch.softmax(attn_map, dim=0)
+    return attn_map
+
+
+def get_net_attn_map(image_size, batch_size=2, instance_or_negative=False, detach=True):
+    idx = 0 if instance_or_negative else 1
+    net_attn_maps = []
+
+    for name, attn_map in attn_maps.items():
+        attn_map = attn_map.cpu() if detach else attn_map
+        attn_map = torch.chunk(attn_map, batch_size)[idx].squeeze()
+        attn_map = upscale(attn_map, image_size)
+        net_attn_maps.append(attn_map)
+
+    net_attn_maps = torch.mean(torch.stack(net_attn_maps, dim=0), dim=0)
+
+    return net_attn_maps
+
+
+def attnmaps2images(net_attn_maps):
+    # total_attn_scores = 0
+    images = []
+
+    for attn_map in net_attn_maps:
+        attn_map = attn_map.cpu().numpy()
+        # total_attn_scores += attn_map.mean().item()
+
+        normalized_attn_map = (attn_map - np.min(attn_map)) / (np.max(attn_map) - np.min(attn_map)) * 255
+        normalized_attn_map = normalized_attn_map.astype(np.uint8)
+        # print("norm: ", normalized_attn_map.shape)
+        image = Image.fromarray(normalized_attn_map)
+
+        # image = fix_save_attn_map(attn_map)
+        images.append(image)
+
+    # print(total_attn_scores)
+    return images
+
+
+def is_torch2_available():
+    return hasattr(F, "scaled_dot_product_attention")

foleycrafter/models/auffusion/attention.py

@@ -0,0 +1,668 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from diffusers.models.lora import LoRACompatibleLinear
+from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
+from diffusers.utils import USE_PEFT_BACKEND
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from .attention_processor import Attention
+
+
+def _chunked_feed_forward(
+    ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None
+):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    if lora_scale is None:
+        ff_output = torch.cat(
+            [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+    else:
+        # TODO(Patrick): LoRA scale can be removed once PEFT refactor is complete
+        ff_output = torch.cat(
+            [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+
+    return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
+        ada_norm_bias: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_continuous:
+            self.norm1 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "rms_norm",
+            )
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if (only_cross_attention and not double_self_attention) else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            if self.use_ada_layer_norm:
+                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            elif self.use_ada_layer_norm_continuous:
+                self.norm2 = AdaLayerNormContinuous(
+                    dim,
+                    ada_norm_continous_conditioning_embedding_dim,
+                    norm_elementwise_affine,
+                    norm_eps,
+                    ada_norm_bias,
+                    "rms_norm",
+                )
+            else:
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if self.use_ada_layer_norm_continuous:
+            self.norm3 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "layer_norm",
+            )
+        elif not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            elif self.use_ada_layer_norm_continuous:
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(
+                self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm_in = nn.LayerNorm(dim)
+        self.ff_in = FeedForward(
+            dim,
+            dim_out=time_mix_inner_dim,
+            activation_fn="geglu",
+        )
+
+        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+        self.attn1 = Attention(
+            query_dim=time_mix_inner_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=None,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+            self.attn2 = Attention(
+                query_dim=time_mix_inner_dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = None
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+        self._chunk_dim = 1
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+        return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        kv_input_dim: int,
+        kv_input_dim_proj_use_bias: bool,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        if kv_input_dim != dim:
+            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+        else:
+            self.kv_mapper = None
+
+        self.norm1 = RMSNorm(dim, 1e-06)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=attention_out_bias,
+        )
+
+        self.norm2 = RMSNorm(dim, 1e-06)
+
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        if self.kv_mapper is not None:
+            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        norm_hidden_states = self.norm2(hidden_states)
+
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim, bias=bias)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out, bias=bias))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+        compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
+        for module in self.net:
+            if isinstance(module, compatible_cls):
+                hidden_states = module(hidden_states, scale)
+            else:
+                hidden_states = module(hidden_states)
+        return hidden_states

foleycrafter/models/auffusion/dual_transformer_2d.py

@@ -0,0 +1,155 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional
+
+from torch import nn
+
+from .transformer_2d import Transformer2DModel, Transformer2DModelOutput
+
+
+class DualTransformer2DModel(nn.Module):
+    """
+    Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            Pass if the input is continuous. The number of channels in the input and output.
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
+        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
+            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
+            `ImagePositionalEmbeddings`.
+        num_vector_embeds (`int`, *optional*):
+            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
+            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
+            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
+            up to but not more than steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the TransformerBlocks' attention should contain a bias parameter.
+    """
+
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+    ):
+        super().__init__()
+        self.transformers = nn.ModuleList(
+            [
+                Transformer2DModel(
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    in_channels=in_channels,
+                    num_layers=num_layers,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    sample_size=sample_size,
+                    num_vector_embeds=num_vector_embeds,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                )
+                for _ in range(2)
+            ]
+        )
+
+        # Variables that can be set by a pipeline:
+
+        # The ratio of transformer1 to transformer2's output states to be combined during inference
+        self.mix_ratio = 0.5
+
+        # The shape of `encoder_hidden_states` is expected to be
+        # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)`
+        self.condition_lengths = [77, 257]
+
+        # Which transformer to use to encode which condition.
+        # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])`
+        self.transformer_index_for_condition = [1, 0]
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states,
+        timestep=None,
+        attention_mask=None,
+        cross_attention_kwargs=None,
+        return_dict: bool = True,
+    ):
+        """
+        Args:
+            hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
+                When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
+                hidden_states.
+            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.long`, *optional*):
+                Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
+            attention_mask (`torch.FloatTensor`, *optional*):
+                Optional attention mask to be applied in Attention.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`:
+            [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        input_states = hidden_states
+
+        encoded_states = []
+        tokens_start = 0
+        # attention_mask is not used yet
+        for i in range(2):
+            # for each of the two transformers, pass the corresponding condition tokens
+            condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]]
+            transformer_index = self.transformer_index_for_condition[i]
+            encoded_state = self.transformers[transformer_index](
+                input_states,
+                encoder_hidden_states=condition_state,
+                timestep=timestep,
+                cross_attention_kwargs=cross_attention_kwargs,
+                return_dict=False,
+            )[0]
+            encoded_states.append(encoded_state - input_states)
+            tokens_start += self.condition_lengths[i]
+
+        output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio)
+        output_states = output_states + input_states
+
+        if not return_dict:
+            return (output_states,)
+
+        return Transformer2DModelOutput(sample=output_states)

foleycrafter/models/auffusion/loaders/ip_adapter.py

@@ -0,0 +1,524 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from pathlib import Path
+from typing import Dict, List, Optional, Union
+
+import torch
+from huggingface_hub.utils import validate_hf_hub_args
+from safetensors import safe_open
+
+from diffusers.models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+from diffusers.utils import (
+    _get_model_file,
+    is_accelerate_available,
+    is_torch_version,
+    is_transformers_available,
+    logging,
+)
+
+
+if is_transformers_available():
+    from transformers import (
+        CLIPImageProcessor,
+        CLIPVisionModelWithProjection,
+    )
+
+    from diffusers.models.attention_processor import (
+        IPAdapterAttnProcessor,
+    )
+
+from ..attention_processor import (
+    IPAdapterAttnProcessor2_0,
+    VPTemporalAdapterAttnProcessor2_0,
+)
+
+
+logger = logging.get_logger(__name__)
+
+
+class IPAdapterMixin:
+    """Mixin for handling IP Adapters."""
+
+    @validate_hf_hub_args
+    def load_ip_adapter(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, List[str], Dict[str, torch.Tensor]],
+        subfolder: Union[str, List[str]],
+        weight_name: Union[str, List[str]],
+        image_encoder_folder: Optional[str] = "image_encoder",
+        **kwargs,
+    ):
+        """
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `List[str]` or `os.PathLike` or `List[os.PathLike]` or `dict` or `List[dict]`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+            subfolder (`str` or `List[str]`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+                If a list is passed, it should have the same length as `weight_name`.
+            weight_name (`str` or `List[str]`):
+                The name of the weight file to load. If a list is passed, it should have the same length as
+                `weight_name`.
+            image_encoder_folder (`str`, *optional*, defaults to `image_encoder`):
+                The subfolder location of the image encoder within a larger model repository on the Hub or locally.
+                Pass `None` to not load the image encoder. If the image encoder is located in a folder inside `subfolder`,
+                you only need to pass the name of the folder that contains image encoder weights, e.g. `image_encoder_folder="image_encoder"`.
+                If the image encoder is located in a folder other than `subfolder`, you should pass the path to the folder that contains image encoder weights,
+                for example, `image_encoder_folder="different_subfolder/image_encoder"`.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+        """
+
+        # handle the list inputs for multiple IP Adapters
+        if not isinstance(weight_name, list):
+            weight_name = [weight_name]
+
+        if not isinstance(pretrained_model_name_or_path_or_dict, list):
+            pretrained_model_name_or_path_or_dict = [pretrained_model_name_or_path_or_dict]
+        if len(pretrained_model_name_or_path_or_dict) == 1:
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict * len(weight_name)
+
+        if not isinstance(subfolder, list):
+            subfolder = [subfolder]
+        if len(subfolder) == 1:
+            subfolder = subfolder * len(weight_name)
+
+        if len(weight_name) != len(pretrained_model_name_or_path_or_dict):
+            raise ValueError("`weight_name` and `pretrained_model_name_or_path_or_dict` must have the same length.")
+
+        if len(weight_name) != len(subfolder):
+            raise ValueError("`weight_name` and `subfolder` must have the same length.")
+
+        # Load the main state dict first.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+
+        if low_cpu_mem_usage and not is_accelerate_available():
+            low_cpu_mem_usage = False
+            logger.warning(
+                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                " install accelerate\n```\n."
+            )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+        state_dicts = []
+        for pretrained_model_name_or_path_or_dict, weight_name, subfolder in zip(
+            pretrained_model_name_or_path_or_dict, weight_name, subfolder
+        ):
+            if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                if weight_name.endswith(".safetensors"):
+                    state_dict = {"image_proj": {}, "ip_adapter": {}}
+                    with safe_open(model_file, framework="pt", device="cpu") as f:
+                        for key in f.keys():
+                            if key.startswith("image_proj."):
+                                state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                            elif key.startswith("ip_adapter."):
+                                state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+                else:
+                    state_dict = torch.load(model_file, map_location="cpu")
+            else:
+                state_dict = pretrained_model_name_or_path_or_dict
+
+            keys = list(state_dict.keys())
+            if keys != ["image_proj", "ip_adapter"]:
+                raise ValueError("Required keys are (`image_proj` and `ip_adapter`) missing from the state dict.")
+
+            state_dicts.append(state_dict)
+
+            # load CLIP image encoder here if it has not been registered to the pipeline yet
+            if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is None:
+                if image_encoder_folder is not None:
+                    if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+                        logger.info(f"loading image_encoder from {pretrained_model_name_or_path_or_dict}")
+                        if image_encoder_folder.count("/") == 0:
+                            image_encoder_subfolder = Path(subfolder, image_encoder_folder).as_posix()
+                        else:
+                            image_encoder_subfolder = Path(image_encoder_folder).as_posix()
+
+                        image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+                            pretrained_model_name_or_path_or_dict,
+                            subfolder=image_encoder_subfolder,
+                            low_cpu_mem_usage=low_cpu_mem_usage,
+                        ).to(self.device, dtype=self.dtype)
+                        self.register_modules(image_encoder=image_encoder)
+                    else:
+                        raise ValueError(
+                            "`image_encoder` cannot be loaded because `pretrained_model_name_or_path_or_dict` is a state dict."
+                        )
+                else:
+                    logger.warning(
+                        "image_encoder is not loaded since `image_encoder_folder=None` passed. You will not be able to use `ip_adapter_image` when calling the pipeline with IP-Adapter."
+                        "Use `ip_adapter_image_embeds` to pass pre-generated image embedding instead."
+                    )
+
+            # create feature extractor if it has not been registered to the pipeline yet
+            if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is None:
+                feature_extractor = CLIPImageProcessor()
+                self.register_modules(feature_extractor=feature_extractor)
+
+        # load ip-adapter into unet
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        unet._load_ip_adapter_weights(state_dicts)
+
+    def set_ip_adapter_scale(self, scale):
+        """
+        Sets the conditioning scale between text and image.
+
+        Example:
+
+        ```py
+        pipeline.set_ip_adapter_scale(0.5)
+        ```
+        """
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        for attn_processor in unet.attn_processors.values():
+            if isinstance(attn_processor, (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0)):
+                if not isinstance(scale, list):
+                    scale = [scale] * len(attn_processor.scale)
+                if len(attn_processor.scale) != len(scale):
+                    raise ValueError(
+                        f"`scale` should be a list of same length as the number if ip-adapters "
+                        f"Expected {len(attn_processor.scale)} but got {len(scale)}."
+                    )
+                attn_processor.scale = scale
+
+    def unload_ip_adapter(self):
+        """
+        Unloads the IP Adapter weights
+
+        Examples:
+
+        ```python
+        >>> # Assuming `pipeline` is already loaded with the IP Adapter weights.
+        >>> pipeline.unload_ip_adapter()
+        >>> ...
+        ```
+        """
+        # remove CLIP image encoder
+        if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is not None:
+            self.image_encoder = None
+            self.register_to_config(image_encoder=[None, None])
+
+        # remove feature extractor only when safety_checker is None as safety_checker uses
+        # the feature_extractor later
+        if not hasattr(self, "safety_checker"):
+            if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is not None:
+                self.feature_extractor = None
+                self.register_to_config(feature_extractor=[None, None])
+
+        # remove hidden encoder
+        self.unet.encoder_hid_proj = None
+        self.config.encoder_hid_dim_type = None
+
+        # restore original Unet attention processors layers
+        self.unet.set_default_attn_processor()
+
+
+class VPAdapterMixin:
+    """Mixin for handling IP Adapters."""
+
+    @validate_hf_hub_args
+    def load_ip_adapter(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, List[str], Dict[str, torch.Tensor]],
+        subfolder: Union[str, List[str]],
+        weight_name: Union[str, List[str]],
+        image_encoder_folder: Optional[str] = "image_encoder",
+        **kwargs,
+    ):
+        """
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `List[str]` or `os.PathLike` or `List[os.PathLike]` or `dict` or `List[dict]`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+            subfolder (`str` or `List[str]`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+                If a list is passed, it should have the same length as `weight_name`.
+            weight_name (`str` or `List[str]`):
+                The name of the weight file to load. If a list is passed, it should have the same length as
+                `weight_name`.
+            image_encoder_folder (`str`, *optional*, defaults to `image_encoder`):
+                The subfolder location of the image encoder within a larger model repository on the Hub or locally.
+                Pass `None` to not load the image encoder. If the image encoder is located in a folder inside `subfolder`,
+                you only need to pass the name of the folder that contains image encoder weights, e.g. `image_encoder_folder="image_encoder"`.
+                If the image encoder is located in a folder other than `subfolder`, you should pass the path to the folder that contains image encoder weights,
+                for example, `image_encoder_folder="different_subfolder/image_encoder"`.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+        """
+
+        # handle the list inputs for multiple IP Adapters
+        if not isinstance(weight_name, list):
+            weight_name = [weight_name]
+
+        if not isinstance(pretrained_model_name_or_path_or_dict, list):
+            pretrained_model_name_or_path_or_dict = [pretrained_model_name_or_path_or_dict]
+        if len(pretrained_model_name_or_path_or_dict) == 1:
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict * len(weight_name)
+
+        if not isinstance(subfolder, list):
+            subfolder = [subfolder]
+        if len(subfolder) == 1:
+            subfolder = subfolder * len(weight_name)
+
+        if len(weight_name) != len(pretrained_model_name_or_path_or_dict):
+            raise ValueError("`weight_name` and `pretrained_model_name_or_path_or_dict` must have the same length.")
+
+        if len(weight_name) != len(subfolder):
+            raise ValueError("`weight_name` and `subfolder` must have the same length.")
+
+        # Load the main state dict first.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+
+        if low_cpu_mem_usage and not is_accelerate_available():
+            low_cpu_mem_usage = False
+            logger.warning(
+                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                " install accelerate\n```\n."
+            )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+        state_dicts = []
+        for pretrained_model_name_or_path_or_dict, weight_name, subfolder in zip(
+            pretrained_model_name_or_path_or_dict, weight_name, subfolder
+        ):
+            if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                if weight_name.endswith(".safetensors"):
+                    state_dict = {"image_proj": {}, "ip_adapter": {}}
+                    with safe_open(model_file, framework="pt", device="cpu") as f:
+                        for key in f.keys():
+                            if key.startswith("image_proj."):
+                                state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                            elif key.startswith("ip_adapter."):
+                                state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+                else:
+                    state_dict = torch.load(model_file, map_location="cpu")
+            else:
+                state_dict = pretrained_model_name_or_path_or_dict
+
+            keys = list(state_dict.keys())
+            if keys != ["image_proj", "ip_adapter"]:
+                raise ValueError("Required keys are (`image_proj` and `ip_adapter`) missing from the state dict.")
+
+            state_dicts.append(state_dict)
+
+            # load CLIP image encoder here if it has not been registered to the pipeline yet
+            if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is None:
+                if image_encoder_folder is not None:
+                    if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+                        logger.info(f"loading image_encoder from {pretrained_model_name_or_path_or_dict}")
+                        if image_encoder_folder.count("/") == 0:
+                            image_encoder_subfolder = Path(subfolder, image_encoder_folder).as_posix()
+                        else:
+                            image_encoder_subfolder = Path(image_encoder_folder).as_posix()
+
+                        image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+                            pretrained_model_name_or_path_or_dict,
+                            subfolder=image_encoder_subfolder,
+                            low_cpu_mem_usage=low_cpu_mem_usage,
+                        ).to(self.device, dtype=self.dtype)
+                        self.register_modules(image_encoder=image_encoder)
+                    else:
+                        raise ValueError(
+                            "`image_encoder` cannot be loaded because `pretrained_model_name_or_path_or_dict` is a state dict."
+                        )
+                else:
+                    logger.warning(
+                        "image_encoder is not loaded since `image_encoder_folder=None` passed. You will not be able to use `ip_adapter_image` when calling the pipeline with IP-Adapter."
+                        "Use `ip_adapter_image_embeds` to pass pre-generated image embedding instead."
+                    )
+
+            # create feature extractor if it has not been registered to the pipeline yet
+            if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is None:
+                feature_extractor = CLIPImageProcessor()
+                self.register_modules(feature_extractor=feature_extractor)
+
+        # load ip-adapter into unet
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        unet._load_ip_adapter_weights_VPAdapter(state_dicts)
+
+    def set_ip_adapter_scale(self, scale):
+        """
+        Sets the conditioning scale between text and image.
+
+        Example:
+
+        ```py
+        pipeline.set_ip_adapter_scale(0.5)
+        ```
+        """
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        for attn_processor in unet.attn_processors.values():
+            if isinstance(attn_processor, (IPAdapterAttnProcessor, VPTemporalAdapterAttnProcessor2_0)):
+                if not isinstance(scale, list):
+                    scale = [scale] * len(attn_processor.scale)
+                if len(attn_processor.scale) != len(scale):
+                    raise ValueError(
+                        f"`scale` should be a list of same length as the number if ip-adapters "
+                        f"Expected {len(attn_processor.scale)} but got {len(scale)}."
+                    )
+                attn_processor.scale = scale
+
+    def unload_ip_adapter(self):
+        """
+        Unloads the IP Adapter weights
+
+        Examples:
+
+        ```python
+        >>> # Assuming `pipeline` is already loaded with the IP Adapter weights.
+        >>> pipeline.unload_ip_adapter()
+        >>> ...
+        ```
+        """
+        # remove CLIP image encoder
+        if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is not None:
+            self.image_encoder = None
+            self.register_to_config(image_encoder=[None, None])
+
+        # remove feature extractor only when safety_checker is None as safety_checker uses
+        # the feature_extractor later
+        if not hasattr(self, "safety_checker"):
+            if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is not None:
+                self.feature_extractor = None
+                self.register_to_config(feature_extractor=[None, None])
+
+        # remove hidden encoder
+        self.unet.encoder_hid_proj = None
+        self.config.encoder_hid_dim_type = None
+
+        # restore original Unet attention processors layers
+        self.unet.set_default_attn_processor()

foleycrafter/models/auffusion/loaders/unet.py

@@ -0,0 +1,992 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from collections import defaultdict
+from contextlib import nullcontext
+from functools import partial
+from typing import Callable, Dict, List, Optional, Tuple, Union
+
+import safetensors
+import torch
+import torch.nn.functional as F
+from huggingface_hub.utils import validate_hf_hub_args
+from torch import nn
+
+from diffusers.loaders.utils import AttnProcsLayers
+from diffusers.models.embeddings import ImageProjection
+from diffusers.models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    _get_model_file,
+    delete_adapter_layers,
+    is_accelerate_available,
+    is_torch_version,
+    logging,
+    set_adapter_layers,
+    set_weights_and_activate_adapters,
+)
+from ..attention_processor import (
+    AttnProcessor2_0,
+    IPAdapterAttnProcessor2_0,
+    VPTemporalAdapterAttnProcessor2_0,
+)
+
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+
+class VPAdapterImageProjection(nn.Module):
+    def __init__(self, IPAdapterImageProjectionLayers: Union[List[nn.Module], Tuple[nn.Module]]):
+        super().__init__()
+        self.image_projection_layers = nn.ModuleList(IPAdapterImageProjectionLayers)
+
+    def forward(self, image_embeds: List[torch.FloatTensor]):
+        projected_image_embeds = []
+
+        # currently, we accept `image_embeds` as
+        #  1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim]
+        #  2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim]
+        if not isinstance(image_embeds, list):
+            # deprecation_message = (
+            #     "You have passed a tensor as `image_embeds`.This is deprecated and will be removed in a future release."
+            #     " Please make sure to update your script to pass `image_embeds` as a list of tensors to suppress this warning."
+            # )
+            image_embeds = [image_embeds.unsqueeze(1)]
+
+        if len(image_embeds) != len(self.image_projection_layers):
+            raise ValueError(
+                f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}"
+            )
+
+        for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers):
+            image_embed = image_embed.squeeze(1)
+            batch_size, num_images = image_embed.shape[0], image_embed.shape[1]
+            image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:])
+            image_embed = image_projection_layer(image_embed)
+            image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:])
+
+            projected_image_embeds.append(image_embed)
+
+        return projected_image_embeds
+
+
+class MultiIPAdapterImageProjection(nn.Module):
+    def __init__(self, IPAdapterImageProjectionLayers: Union[List[nn.Module], Tuple[nn.Module]]):
+        super().__init__()
+        self.image_projection_layers = nn.ModuleList(IPAdapterImageProjectionLayers)
+
+    def forward(self, image_embeds: List[torch.FloatTensor]):
+        projected_image_embeds = []
+
+        # currently, we accept `image_embeds` as
+        #  1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim]
+        #  2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim]
+        if not isinstance(image_embeds, list):
+            # deprecation_message = (
+            #     "You have passed a tensor as `image_embeds`.This is deprecated and will be removed in a future release."
+            #     " Please make sure to update your script to pass `image_embeds` as a list of tensors to suppress this warning."
+            # )
+            image_embeds = [image_embeds.unsqueeze(1)]
+
+        if len(image_embeds) != len(self.image_projection_layers):
+            raise ValueError(
+                f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}"
+            )
+
+        for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers):
+            batch_size, num_images = image_embed.shape[0], image_embed.shape[1]
+            image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:])
+            image_embed = image_projection_layer(image_embed)
+            image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:])
+
+            projected_image_embeds.append(image_embed)
+
+        return projected_image_embeds
+
+
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+CUSTOM_DIFFUSION_WEIGHT_NAME = "pytorch_custom_diffusion_weights.bin"
+CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "pytorch_custom_diffusion_weights.safetensors"
+
+
+class UNet2DConditionLoadersMixin:
+    """
+    Load LoRA layers into a [`UNet2DCondtionModel`].
+    """
+
+    text_encoder_name = TEXT_ENCODER_NAME
+    unet_name = UNET_NAME
+
+    @validate_hf_hub_args
+    def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
+        r"""
+        Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
+        defined in
+        [`attention_processor.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)
+        and be a `torch.nn.Module` class.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the model id (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a directory (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.unet.load_attn_procs(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        ```
+        """
+        from diffusers.models.attention_processor import CustomDiffusionAttnProcessor
+        from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
+
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
+        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+        network_alphas = kwargs.pop("network_alphas", None)
+
+        _pipeline = kwargs.pop("_pipeline", None)
+
+        is_network_alphas_none = network_alphas is None
+
+        allow_pickle = False
+
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        if low_cpu_mem_usage and not is_accelerate_available():
+            low_cpu_mem_usage = False
+            logger.warning(
+                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                " install accelerate\n```\n."
+            )
+
+        model_file = None
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        token=token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                    )
+                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
+                except IOError as e:
+                    if not allow_pickle:
+                        raise e
+                    # try loading non-safetensors weights
+                    pass
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name or LORA_WEIGHT_NAME,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        # fill attn processors
+        lora_layers_list = []
+
+        is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys()) and not USE_PEFT_BACKEND
+        is_custom_diffusion = any("custom_diffusion" in k for k in state_dict.keys())
+
+        if is_lora:
+            # correct keys
+            state_dict, network_alphas = self.convert_state_dict_legacy_attn_format(state_dict, network_alphas)
+
+            if network_alphas is not None:
+                network_alphas_keys = list(network_alphas.keys())
+                used_network_alphas_keys = set()
+
+            lora_grouped_dict = defaultdict(dict)
+            mapped_network_alphas = {}
+
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                value = state_dict.pop(key)
+                attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                lora_grouped_dict[attn_processor_key][sub_key] = value
+
+                # Create another `mapped_network_alphas` dictionary so that we can properly map them.
+                if network_alphas is not None:
+                    for k in network_alphas_keys:
+                        if k.replace(".alpha", "") in key:
+                            mapped_network_alphas.update({attn_processor_key: network_alphas.get(k)})
+                            used_network_alphas_keys.add(k)
+
+            if not is_network_alphas_none:
+                if len(set(network_alphas_keys) - used_network_alphas_keys) > 0:
+                    raise ValueError(
+                        f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
+                    )
+
+            if len(state_dict) > 0:
+                raise ValueError(
+                    f"The `state_dict` has to be empty at this point but has the following keys \n\n {', '.join(state_dict.keys())}"
+                )
+
+            for key, value_dict in lora_grouped_dict.items():
+                attn_processor = self
+                for sub_key in key.split("."):
+                    attn_processor = getattr(attn_processor, sub_key)
+
+                # Process non-attention layers, which don't have to_{k,v,q,out_proj}_lora layers
+                # or add_{k,v,q,out_proj}_proj_lora layers.
+                rank = value_dict["lora.down.weight"].shape[0]
+
+                if isinstance(attn_processor, LoRACompatibleConv):
+                    in_features = attn_processor.in_channels
+                    out_features = attn_processor.out_channels
+                    kernel_size = attn_processor.kernel_size
+
+                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRAConv2dLayer(
+                            in_features=in_features,
+                            out_features=out_features,
+                            rank=rank,
+                            kernel_size=kernel_size,
+                            stride=attn_processor.stride,
+                            padding=attn_processor.padding,
+                            network_alpha=mapped_network_alphas.get(key),
+                        )
+                elif isinstance(attn_processor, LoRACompatibleLinear):
+                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRALinearLayer(
+                            attn_processor.in_features,
+                            attn_processor.out_features,
+                            rank,
+                            mapped_network_alphas.get(key),
+                        )
+                else:
+                    raise ValueError(f"Module {key} is not a LoRACompatibleConv or LoRACompatibleLinear module.")
+
+                value_dict = {k.replace("lora.", ""): v for k, v in value_dict.items()}
+                lora_layers_list.append((attn_processor, lora))
+
+                if low_cpu_mem_usage:
+                    device = next(iter(value_dict.values())).device
+                    dtype = next(iter(value_dict.values())).dtype
+                    load_model_dict_into_meta(lora, value_dict, device=device, dtype=dtype)
+                else:
+                    lora.load_state_dict(value_dict)
+
+        elif is_custom_diffusion:
+            attn_processors = {}
+            custom_diffusion_grouped_dict = defaultdict(dict)
+            for key, value in state_dict.items():
+                if len(value) == 0:
+                    custom_diffusion_grouped_dict[key] = {}
+                else:
+                    if "to_out" in key:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                    else:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
+                    custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
+
+            for key, value_dict in custom_diffusion_grouped_dict.items():
+                if len(value_dict) == 0:
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
+                    )
+                else:
+                    cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
+                    hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
+                    train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=True,
+                        train_q_out=train_q_out,
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                    )
+                    attn_processors[key].load_state_dict(value_dict)
+        elif USE_PEFT_BACKEND:
+            # In that case we have nothing to do as loading the adapter weights is already handled above by `set_peft_model_state_dict`
+            # on the Unet
+            pass
+        else:
+            raise ValueError(
+                f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
+            )
+
+        # <Unsafe code
+        # We can be sure that the following works as it just sets attention processors, lora layers and puts all in the same dtype
+        # Now we remove any existing hooks to
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+
+        # For PEFT backend the Unet is already offloaded at this stage as it is handled inside `lora_lora_weights_into_unet`
+        if not USE_PEFT_BACKEND:
+            if _pipeline is not None:
+                for _, component in _pipeline.components.items():
+                    if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
+                        is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                        is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+
+                        logger.info(
+                            "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                        )
+                        remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+            # only custom diffusion needs to set attn processors
+            if is_custom_diffusion:
+                self.set_attn_processor(attn_processors)
+
+            # set lora layers
+            for target_module, lora_layer in lora_layers_list:
+                target_module.set_lora_layer(lora_layer)
+
+            self.to(dtype=self.dtype, device=self.device)
+
+            # Offload back.
+            if is_model_cpu_offload:
+                _pipeline.enable_model_cpu_offload()
+            elif is_sequential_cpu_offload:
+                _pipeline.enable_sequential_cpu_offload()
+            # Unsafe code />
+
+    def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
+        is_new_lora_format = all(
+            key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
+        )
+        if is_new_lora_format:
+            # Strip the `"unet"` prefix.
+            is_text_encoder_present = any(key.startswith(self.text_encoder_name) for key in state_dict.keys())
+            if is_text_encoder_present:
+                warn_message = "The state_dict contains LoRA params corresponding to the text encoder which are not being used here. To use both UNet and text encoder related LoRA params, use [`pipe.load_lora_weights()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights)."
+                logger.warn(warn_message)
+            unet_keys = [k for k in state_dict.keys() if k.startswith(self.unet_name)]
+            state_dict = {k.replace(f"{self.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+        # change processor format to 'pure' LoRACompatibleLinear format
+        if any("processor" in k.split(".") for k in state_dict.keys()):
+
+            def format_to_lora_compatible(key):
+                if "processor" not in key.split("."):
+                    return key
+                return key.replace(".processor", "").replace("to_out_lora", "to_out.0.lora").replace("_lora", ".lora")
+
+            state_dict = {format_to_lora_compatible(k): v for k, v in state_dict.items()}
+
+            if network_alphas is not None:
+                network_alphas = {format_to_lora_compatible(k): v for k, v in network_alphas.items()}
+        return state_dict, network_alphas
+
+    def save_attn_procs(
+        self,
+        save_directory: Union[str, os.PathLike],
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        **kwargs,
+    ):
+        r"""
+        Save attention processor layers to a directory so that it can be reloaded with the
+        [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save an attention processor to (will be created if it doesn't exist).
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or with `pickle`.
+
+        Example:
+
+        ```py
+        import torch
+        from diffusers import DiffusionPipeline
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "CompVis/stable-diffusion-v1-4",
+            torch_dtype=torch.float16,
+        ).to("cuda")
+        pipeline.unet.load_attn_procs("path-to-save-model", weight_name="pytorch_custom_diffusion_weights.bin")
+        pipeline.unet.save_attn_procs("path-to-save-model", weight_name="pytorch_custom_diffusion_weights.bin")
+        ```
+        """
+        from diffusers.models.attention_processor import (
+            CustomDiffusionAttnProcessor,
+            CustomDiffusionAttnProcessor2_0,
+            CustomDiffusionXFormersAttnProcessor,
+        )
+
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        if save_function is None:
+            if safe_serialization:
+
+                def save_function(weights, filename):
+                    return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
+
+            else:
+                save_function = torch.save
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        is_custom_diffusion = any(
+            isinstance(
+                x,
+                (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor),
+            )
+            for (_, x) in self.attn_processors.items()
+        )
+        if is_custom_diffusion:
+            model_to_save = AttnProcsLayers(
+                {
+                    y: x
+                    for (y, x) in self.attn_processors.items()
+                    if isinstance(
+                        x,
+                        (
+                            CustomDiffusionAttnProcessor,
+                            CustomDiffusionAttnProcessor2_0,
+                            CustomDiffusionXFormersAttnProcessor,
+                        ),
+                    )
+                }
+            )
+            state_dict = model_to_save.state_dict()
+            for name, attn in self.attn_processors.items():
+                if len(attn.state_dict()) == 0:
+                    state_dict[name] = {}
+        else:
+            model_to_save = AttnProcsLayers(self.attn_processors)
+            state_dict = model_to_save.state_dict()
+
+        if weight_name is None:
+            if safe_serialization:
+                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE if is_custom_diffusion else LORA_WEIGHT_NAME_SAFE
+            else:
+                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME if is_custom_diffusion else LORA_WEIGHT_NAME
+
+        # Save the model
+        save_function(state_dict, os.path.join(save_directory, weight_name))
+        logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
+
+    def fuse_lora(self, lora_scale=1.0, safe_fusing=False, adapter_names=None):
+        self.lora_scale = lora_scale
+        self._safe_fusing = safe_fusing
+        self.apply(partial(self._fuse_lora_apply, adapter_names=adapter_names))
+
+    def _fuse_lora_apply(self, module, adapter_names=None):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_fuse_lora"):
+                module._fuse_lora(self.lora_scale, self._safe_fusing)
+
+            if adapter_names is not None:
+                raise ValueError(
+                    "The `adapter_names` argument is not supported in your environment. Please switch"
+                    " to PEFT backend to use this argument by installing latest PEFT and transformers."
+                    " `pip install -U peft transformers`"
+                )
+
+    def unfuse_lora(self):
+        self.apply(self._unfuse_lora_apply)
+
+    def _unfuse_lora_apply(self, module):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_unfuse_lora"):
+                module._unfuse_lora()
+
+    def set_adapters(
+        self,
+        adapter_names: Union[List[str], str],
+        weights: Optional[Union[List[float], float]] = None,
+    ):
+        """
+        Set the currently active adapters for use in the UNet.
+
+        Args:
+            adapter_names (`List[str]` or `str`):
+                The names of the adapters to use.
+            adapter_weights (`Union[List[float], float]`, *optional*):
+                The adapter(s) weights to use with the UNet. If `None`, the weights are set to `1.0` for all the
+                adapters.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.set_adapters(["cinematic", "pixel"], adapter_weights=[0.5, 0.5])
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for `set_adapters()`.")
+
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+        if weights is None:
+            weights = [1.0] * len(adapter_names)
+        elif isinstance(weights, float):
+            weights = [weights] * len(adapter_names)
+
+        if len(adapter_names) != len(weights):
+            raise ValueError(
+                f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
+            )
+
+        set_weights_and_activate_adapters(self, adapter_names, weights)
+
+    def disable_lora(self):
+        """
+        Disable the UNet's active LoRA layers.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.disable_lora()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=False)
+
+    def enable_lora(self):
+        """
+        Enable the UNet's active LoRA layers.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.enable_lora()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=True)
+
+    def delete_adapters(self, adapter_names: Union[List[str], str]):
+        """
+        Delete an adapter's LoRA layers from the UNet.
+
+        Args:
+            adapter_names (`Union[List[str], str]`):
+                The names (single string or list of strings) of the adapter to delete.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_names="cinematic"
+        )
+        pipeline.delete_adapters("cinematic")
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        if isinstance(adapter_names, str):
+            adapter_names = [adapter_names]
+
+        for adapter_name in adapter_names:
+            delete_adapter_layers(self, adapter_name)
+
+            # Pop also the corresponding adapter from the config
+            if hasattr(self, "peft_config"):
+                self.peft_config.pop(adapter_name, None)
+
+    def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_usage=False):
+        if low_cpu_mem_usage:
+            if is_accelerate_available():
+                from accelerate import init_empty_weights
+
+            else:
+                low_cpu_mem_usage = False
+                logger.warning(
+                    "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                    " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                    " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                    " install accelerate\n```\n."
+                )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        updated_state_dict = {}
+        image_projection = None
+        init_context = init_empty_weights if low_cpu_mem_usage else nullcontext
+
+        if "proj.weight" in state_dict:
+            # IP-Adapter
+            num_image_text_embeds = 4
+            clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+            cross_attention_dim = state_dict["proj.weight"].shape[0] // num_image_text_embeds
+
+            with init_context():
+                image_projection = ImageProjection(
+                    cross_attention_dim=cross_attention_dim,
+                    image_embed_dim=clip_embeddings_dim,
+                    num_image_text_embeds=num_image_text_embeds,
+                )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj", "image_embeds")
+                updated_state_dict[diffusers_name] = value
+
+        if not low_cpu_mem_usage:
+            image_projection.load_state_dict(updated_state_dict)
+        else:
+            load_model_dict_into_meta(image_projection, updated_state_dict, device=self.device, dtype=self.dtype)
+
+        return image_projection
+
+    def _convert_ip_adapter_attn_to_diffusers_VPAdapter(self, state_dicts, low_cpu_mem_usage=False):
+        from diffusers.models.attention_processor import (
+            AttnProcessor,
+            IPAdapterAttnProcessor,
+        )
+
+        if low_cpu_mem_usage:
+            if is_accelerate_available():
+                from accelerate import init_empty_weights
+
+            else:
+                low_cpu_mem_usage = False
+                logger.warning(
+                    "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                    " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                    " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                    " install accelerate\n```\n."
+                )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        # set ip-adapter cross-attention processors & load state_dict
+        attn_procs = {}
+        key_id = 1
+        init_context = init_empty_weights if low_cpu_mem_usage else nullcontext
+        for name in self.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else self.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = self.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(self.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = self.config.block_out_channels[block_id]
+
+            if cross_attention_dim is None or "motion_modules" in name or "fuser" in name:
+                attn_processor_class = (
+                    AttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else AttnProcessor
+                )
+                attn_procs[name] = attn_processor_class()
+            else:
+                attn_processor_class = (
+                    VPTemporalAdapterAttnProcessor2_0
+                    if hasattr(F, "scaled_dot_product_attention")
+                    else IPAdapterAttnProcessor
+                )
+                num_image_text_embeds = []
+                for state_dict in state_dicts:
+                    if "proj.weight" in state_dict["image_proj"]:
+                        # IP-Adapter
+                        num_image_text_embeds += [4]
+                    elif "proj.3.weight" in state_dict["image_proj"]:
+                        # IP-Adapter Full Face
+                        num_image_text_embeds += [257]  # 256 CLIP tokens + 1 CLS token
+                    else:
+                        # IP-Adapter Plus
+                        num_image_text_embeds += [state_dict["image_proj"]["latents"].shape[1]]
+
+                with init_context():
+                    attn_procs[name] = attn_processor_class(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=num_image_text_embeds,
+                    )
+
+                value_dict = {}
+                for i, state_dict in enumerate(state_dicts):
+                    value_dict.update({f"to_k_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_k_ip.weight"]})
+                    value_dict.update({f"to_v_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_v_ip.weight"]})
+
+                if not low_cpu_mem_usage:
+                    attn_procs[name].load_state_dict(value_dict)
+                else:
+                    device = next(iter(value_dict.values())).device
+                    dtype = next(iter(value_dict.values())).dtype
+                    load_model_dict_into_meta(attn_procs[name], value_dict, device=device, dtype=dtype)
+
+                key_id += 2
+
+        return attn_procs
+
+    def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=False):
+        from diffusers.models.attention_processor import (
+            AttnProcessor,
+            IPAdapterAttnProcessor,
+        )
+
+        if low_cpu_mem_usage:
+            if is_accelerate_available():
+                from accelerate import init_empty_weights
+
+            else:
+                low_cpu_mem_usage = False
+                logger.warning(
+                    "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                    " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                    " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                    " install accelerate\n```\n."
+                )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        # set ip-adapter cross-attention processors & load state_dict
+        attn_procs = {}
+        key_id = 1
+        init_context = init_empty_weights if low_cpu_mem_usage else nullcontext
+        for name in self.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else self.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = self.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(self.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = self.config.block_out_channels[block_id]
+
+            if cross_attention_dim is None or "motion_modules" in name or "fuser" in name:
+                attn_processor_class = (
+                    AttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else AttnProcessor
+                )
+                attn_procs[name] = attn_processor_class()
+            else:
+                attn_processor_class = (
+                    IPAdapterAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else IPAdapterAttnProcessor
+                )
+                num_image_text_embeds = []
+                for state_dict in state_dicts:
+                    if "proj.weight" in state_dict["image_proj"]:
+                        # IP-Adapter
+                        num_image_text_embeds += [4]
+                    elif "proj.3.weight" in state_dict["image_proj"]:
+                        # IP-Adapter Full Face
+                        num_image_text_embeds += [257]  # 256 CLIP tokens + 1 CLS token
+                    else:
+                        # IP-Adapter Plus
+                        num_image_text_embeds += [state_dict["image_proj"]["latents"].shape[1]]
+
+                with init_context():
+                    attn_procs[name] = attn_processor_class(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=num_image_text_embeds,
+                    )
+
+                value_dict = {}
+                for i, state_dict in enumerate(state_dicts):
+                    value_dict.update({f"to_k_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_k_ip.weight"]})
+                    value_dict.update({f"to_v_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_v_ip.weight"]})
+
+                if not low_cpu_mem_usage:
+                    attn_procs[name].load_state_dict(value_dict)
+                else:
+                    device = next(iter(value_dict.values())).device
+                    dtype = next(iter(value_dict.values())).dtype
+                    load_model_dict_into_meta(attn_procs[name], value_dict, device=device, dtype=dtype)
+
+                key_id += 2
+
+        return attn_procs
+
+    def _load_ip_adapter_weights(self, state_dicts, low_cpu_mem_usage=False):
+        attn_procs = self._convert_ip_adapter_attn_to_diffusers(state_dicts, low_cpu_mem_usage=low_cpu_mem_usage)
+        self.set_attn_processor(attn_procs)
+
+        # convert IP-Adapter Image Projection layers to diffusers
+        image_projection_layers = []
+        for state_dict in state_dicts:
+            image_projection_layer = self._convert_ip_adapter_image_proj_to_diffusers(
+                state_dict["image_proj"], low_cpu_mem_usage=low_cpu_mem_usage
+            )
+            image_projection_layers.append(image_projection_layer)
+
+        self.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
+        self.config.encoder_hid_dim_type = "ip_image_proj"
+
+        self.to(dtype=self.dtype, device=self.device)
+
+    def _load_ip_adapter_weights_VPAdapter(self, state_dicts, low_cpu_mem_usage=False):
+        attn_procs = self._convert_ip_adapter_attn_to_diffusers_VPAdapter(
+            state_dicts, low_cpu_mem_usage=low_cpu_mem_usage
+        )
+        self.set_attn_processor(attn_procs)
+
+        # convert IP-Adapter Image Projection layers to diffusers
+        image_projection_layers = []
+        for state_dict in state_dicts:
+            image_projection_layer = self._convert_ip_adapter_image_proj_to_diffusers(
+                state_dict["image_proj"], low_cpu_mem_usage=low_cpu_mem_usage
+            )
+            image_projection_layers.append(image_projection_layer)
+
+        self.encoder_hid_proj = VPAdapterImageProjection(image_projection_layers)
+        self.config.encoder_hid_dim_type = "ip_image_proj"
+
+        self.to(dtype=self.dtype, device=self.device)

foleycrafter/models/auffusion/resnet.py

@@ -0,0 +1,685 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+# `TemporalConvLayer` Copyright 2023 Alibaba DAMO-VILAB, The ModelScope Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partial
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.models.activations import get_activation
+from diffusers.models.downsampling import (  # noqa
+    Downsample1D,
+    Downsample2D,
+    FirDownsample2D,
+    KDownsample2D,
+    downsample_2d,
+)
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.normalization import AdaGroupNorm
+from diffusers.models.upsampling import (  # noqa
+    FirUpsample2D,
+    KUpsample2D,
+    Upsample1D,
+    Upsample2D,
+    upfirdn2d_native,
+    upsample_2d,
+)
+from diffusers.utils import USE_PEFT_BACKEND
+from .attention_processor import SpatialNorm
+
+
+class ResnetBlock2D(nn.Module):
+    r"""
+    A Resnet block.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+        dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+        temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+        groups (`int`, *optional*, default to `32`): The number of groups to use for the first normalization layer.
+        groups_out (`int`, *optional*, default to None):
+            The number of groups to use for the second normalization layer. if set to None, same as `groups`.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+        non_linearity (`str`, *optional*, default to `"swish"`): the activation function to use.
+        time_embedding_norm (`str`, *optional*, default to `"default"` ): Time scale shift config.
+            By default, apply timestep embedding conditioning with a simple shift mechanism. Choose "scale_shift" or
+            "ada_group" for a stronger conditioning with scale and shift.
+        kernel (`torch.FloatTensor`, optional, default to None): FIR filter, see
+            [`~models.resnet.FirUpsample2D`] and [`~models.resnet.FirDownsample2D`].
+        output_scale_factor (`float`, *optional*, default to be `1.0`): the scale factor to use for the output.
+        use_in_shortcut (`bool`, *optional*, default to `True`):
+            If `True`, add a 1x1 nn.conv2d layer for skip-connection.
+        up (`bool`, *optional*, default to `False`): If `True`, add an upsample layer.
+        down (`bool`, *optional*, default to `False`): If `True`, add a downsample layer.
+        conv_shortcut_bias (`bool`, *optional*, default to `True`):  If `True`, adds a learnable bias to the
+            `conv_shortcut` output.
+        conv_2d_out_channels (`int`, *optional*, default to `None`): the number of channels in the output.
+            If None, same as `out_channels`.
+    """
+
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        conv_shortcut: bool = False,
+        dropout: float = 0.0,
+        temb_channels: int = 512,
+        groups: int = 32,
+        groups_out: Optional[int] = None,
+        pre_norm: bool = True,
+        eps: float = 1e-6,
+        non_linearity: str = "swish",
+        skip_time_act: bool = False,
+        time_embedding_norm: str = "default",  # default, scale_shift, ada_group, spatial
+        kernel: Optional[torch.FloatTensor] = None,
+        output_scale_factor: float = 1.0,
+        use_in_shortcut: Optional[bool] = None,
+        up: bool = False,
+        down: bool = False,
+        conv_shortcut_bias: bool = True,
+        conv_2d_out_channels: Optional[int] = None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.up = up
+        self.down = down
+        self.output_scale_factor = output_scale_factor
+        self.time_embedding_norm = time_embedding_norm
+        self.skip_time_act = skip_time_act
+
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+
+        if groups_out is None:
+            groups_out = groups
+
+        if self.time_embedding_norm == "ada_group":
+            self.norm1 = AdaGroupNorm(temb_channels, in_channels, groups, eps=eps)
+        elif self.time_embedding_norm == "spatial":
+            self.norm1 = SpatialNorm(in_channels, temb_channels)
+        else:
+            self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+
+        self.conv1 = conv_cls(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                self.time_emb_proj = linear_cls(temb_channels, out_channels)
+            elif self.time_embedding_norm == "scale_shift":
+                self.time_emb_proj = linear_cls(temb_channels, 2 * out_channels)
+            elif self.time_embedding_norm == "ada_group" or self.time_embedding_norm == "spatial":
+                self.time_emb_proj = None
+            else:
+                raise ValueError(f"unknown time_embedding_norm : {self.time_embedding_norm} ")
+        else:
+            self.time_emb_proj = None
+
+        if self.time_embedding_norm == "ada_group":
+            self.norm2 = AdaGroupNorm(temb_channels, out_channels, groups_out, eps=eps)
+        elif self.time_embedding_norm == "spatial":
+            self.norm2 = SpatialNorm(out_channels, temb_channels)
+        else:
+            self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+
+        self.dropout = torch.nn.Dropout(dropout)
+        conv_2d_out_channels = conv_2d_out_channels or out_channels
+        self.conv2 = conv_cls(out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1)
+
+        self.nonlinearity = get_activation(non_linearity)
+
+        self.upsample = self.downsample = None
+        if self.up:
+            if kernel == "fir":
+                fir_kernel = (1, 3, 3, 1)
+                self.upsample = lambda x: upsample_2d(x, kernel=fir_kernel)
+            elif kernel == "sde_vp":
+                self.upsample = partial(F.interpolate, scale_factor=2.0, mode="nearest")
+            else:
+                self.upsample = Upsample2D(in_channels, use_conv=False)
+        elif self.down:
+            if kernel == "fir":
+                fir_kernel = (1, 3, 3, 1)
+                self.downsample = lambda x: downsample_2d(x, kernel=fir_kernel)
+            elif kernel == "sde_vp":
+                self.downsample = partial(F.avg_pool2d, kernel_size=2, stride=2)
+            else:
+                self.downsample = Downsample2D(in_channels, use_conv=False, padding=1, name="op")
+
+        self.use_in_shortcut = self.in_channels != conv_2d_out_channels if use_in_shortcut is None else use_in_shortcut
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = conv_cls(
+                in_channels,
+                conv_2d_out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=conv_shortcut_bias,
+            )
+
+    def forward(
+        self,
+        input_tensor: torch.FloatTensor,
+        temb: torch.FloatTensor,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        hidden_states = input_tensor
+
+        if self.time_embedding_norm == "ada_group" or self.time_embedding_norm == "spatial":
+            hidden_states = self.norm1(hidden_states, temb)
+        else:
+            hidden_states = self.norm1(hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        if self.upsample is not None:
+            # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+            if hidden_states.shape[0] >= 64:
+                input_tensor = input_tensor.contiguous()
+                hidden_states = hidden_states.contiguous()
+            input_tensor = (
+                self.upsample(input_tensor, scale=scale)
+                if isinstance(self.upsample, Upsample2D)
+                else self.upsample(input_tensor)
+            )
+            hidden_states = (
+                self.upsample(hidden_states, scale=scale)
+                if isinstance(self.upsample, Upsample2D)
+                else self.upsample(hidden_states)
+            )
+        elif self.downsample is not None:
+            input_tensor = (
+                self.downsample(input_tensor, scale=scale)
+                if isinstance(self.downsample, Downsample2D)
+                else self.downsample(input_tensor)
+            )
+            hidden_states = (
+                self.downsample(hidden_states, scale=scale)
+                if isinstance(self.downsample, Downsample2D)
+                else self.downsample(hidden_states)
+            )
+
+        hidden_states = self.conv1(hidden_states, scale) if not USE_PEFT_BACKEND else self.conv1(hidden_states)
+
+        if self.time_emb_proj is not None:
+            if not self.skip_time_act:
+                temb = self.nonlinearity(temb)
+            temb = (
+                self.time_emb_proj(temb, scale)[:, :, None, None]
+                if not USE_PEFT_BACKEND
+                # NOTE: Maybe we can use different prompt in different time
+                else self.time_emb_proj(temb)[:, :, None, None]
+            )
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        if self.time_embedding_norm == "ada_group" or self.time_embedding_norm == "spatial":
+            hidden_states = self.norm2(hidden_states, temb)
+        else:
+            hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states, scale) if not USE_PEFT_BACKEND else self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = (
+                self.conv_shortcut(input_tensor, scale) if not USE_PEFT_BACKEND else self.conv_shortcut(input_tensor)
+            )
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+# unet_rl.py
+def rearrange_dims(tensor: torch.Tensor) -> torch.Tensor:
+    if len(tensor.shape) == 2:
+        return tensor[:, :, None]
+    if len(tensor.shape) == 3:
+        return tensor[:, :, None, :]
+    elif len(tensor.shape) == 4:
+        return tensor[:, :, 0, :]
+    else:
+        raise ValueError(f"`len(tensor)`: {len(tensor)} has to be 2, 3 or 4.")
+
+
+class Conv1dBlock(nn.Module):
+    """
+    Conv1d --> GroupNorm --> Mish
+
+    Parameters:
+        inp_channels (`int`): Number of input channels.
+        out_channels (`int`): Number of output channels.
+        kernel_size (`int` or `tuple`): Size of the convolving kernel.
+        n_groups (`int`, default `8`): Number of groups to separate the channels into.
+        activation (`str`, defaults to `mish`): Name of the activation function.
+    """
+
+    def __init__(
+        self,
+        inp_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, Tuple[int, int]],
+        n_groups: int = 8,
+        activation: str = "mish",
+    ):
+        super().__init__()
+
+        self.conv1d = nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2)
+        self.group_norm = nn.GroupNorm(n_groups, out_channels)
+        self.mish = get_activation(activation)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        intermediate_repr = self.conv1d(inputs)
+        intermediate_repr = rearrange_dims(intermediate_repr)
+        intermediate_repr = self.group_norm(intermediate_repr)
+        intermediate_repr = rearrange_dims(intermediate_repr)
+        output = self.mish(intermediate_repr)
+        return output
+
+
+# unet_rl.py
+class ResidualTemporalBlock1D(nn.Module):
+    """
+    Residual 1D block with temporal convolutions.
+
+    Parameters:
+        inp_channels (`int`): Number of input channels.
+        out_channels (`int`): Number of output channels.
+        embed_dim (`int`): Embedding dimension.
+        kernel_size (`int` or `tuple`): Size of the convolving kernel.
+        activation (`str`, defaults `mish`): It is possible to choose the right activation function.
+    """
+
+    def __init__(
+        self,
+        inp_channels: int,
+        out_channels: int,
+        embed_dim: int,
+        kernel_size: Union[int, Tuple[int, int]] = 5,
+        activation: str = "mish",
+    ):
+        super().__init__()
+        self.conv_in = Conv1dBlock(inp_channels, out_channels, kernel_size)
+        self.conv_out = Conv1dBlock(out_channels, out_channels, kernel_size)
+
+        self.time_emb_act = get_activation(activation)
+        self.time_emb = nn.Linear(embed_dim, out_channels)
+
+        self.residual_conv = (
+            nn.Conv1d(inp_channels, out_channels, 1) if inp_channels != out_channels else nn.Identity()
+        )
+
+    def forward(self, inputs: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            inputs : [ batch_size x inp_channels x horizon ]
+            t : [ batch_size x embed_dim ]
+
+        returns:
+            out : [ batch_size x out_channels x horizon ]
+        """
+        t = self.time_emb_act(t)
+        t = self.time_emb(t)
+        out = self.conv_in(inputs) + rearrange_dims(t)
+        out = self.conv_out(out)
+        return out + self.residual_conv(inputs)
+
+
+class TemporalConvLayer(nn.Module):
+    """
+    Temporal convolutional layer that can be used for video (sequence of images) input Code mostly copied from:
+    https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/models/multi_modal/video_synthesis/unet_sd.py#L1016
+
+    Parameters:
+        in_dim (`int`): Number of input channels.
+        out_dim (`int`): Number of output channels.
+        dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: Optional[int] = None,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+    ):
+        super().__init__()
+        out_dim = out_dim or in_dim
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # conv layers
+        self.conv1 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, in_dim),
+            nn.SiLU(),
+            nn.Conv3d(in_dim, out_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+        self.conv2 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, out_dim),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+        self.conv3 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, out_dim),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+        self.conv4 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, out_dim),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+
+        # zero out the last layer params,so the conv block is identity
+        nn.init.zeros_(self.conv4[-1].weight)
+        nn.init.zeros_(self.conv4[-1].bias)
+
+    def forward(self, hidden_states: torch.Tensor, num_frames: int = 1) -> torch.Tensor:
+        hidden_states = (
+            hidden_states[None, :].reshape((-1, num_frames) + hidden_states.shape[1:]).permute(0, 2, 1, 3, 4)
+        )
+
+        identity = hidden_states
+        hidden_states = self.conv1(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+        hidden_states = self.conv3(hidden_states)
+        hidden_states = self.conv4(hidden_states)
+
+        hidden_states = identity + hidden_states
+
+        hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(
+            (hidden_states.shape[0] * hidden_states.shape[2], -1) + hidden_states.shape[3:]
+        )
+        return hidden_states
+
+
+class TemporalResnetBlock(nn.Module):
+    r"""
+    A Resnet block.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+        temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        temb_channels: int = 512,
+        eps: float = 1e-6,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+
+        kernel_size = (3, 1, 1)
+        padding = [k // 2 for k in kernel_size]
+
+        self.norm1 = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=eps, affine=True)
+        self.conv1 = nn.Conv3d(
+            in_channels,
+            out_channels,
+            kernel_size=kernel_size,
+            stride=1,
+            padding=padding,
+        )
+
+        if temb_channels is not None:
+            self.time_emb_proj = nn.Linear(temb_channels, out_channels)
+        else:
+            self.time_emb_proj = None
+
+        self.norm2 = torch.nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=eps, affine=True)
+
+        self.dropout = torch.nn.Dropout(0.0)
+        self.conv2 = nn.Conv3d(
+            out_channels,
+            out_channels,
+            kernel_size=kernel_size,
+            stride=1,
+            padding=padding,
+        )
+
+        self.nonlinearity = get_activation("silu")
+
+        self.use_in_shortcut = self.in_channels != out_channels
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = nn.Conv3d(
+                in_channels,
+                out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            )
+
+    def forward(self, input_tensor: torch.FloatTensor, temb: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv1(hidden_states)
+
+        if self.time_emb_proj is not None:
+            temb = self.nonlinearity(temb)
+            temb = self.time_emb_proj(temb)[:, :, :, None, None]
+            temb = temb.permute(0, 2, 1, 3, 4)
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = input_tensor + hidden_states
+
+        return output_tensor
+
+
+# VideoResBlock
+class SpatioTemporalResBlock(nn.Module):
+    r"""
+    A SpatioTemporal Resnet block.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+        temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the spatial resenet.
+        temporal_eps (`float`, *optional*, defaults to `eps`): The epsilon to use for the temporal resnet.
+        merge_factor (`float`, *optional*, defaults to `0.5`): The merge factor to use for the temporal mixing.
+        merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
+            The merge strategy to use for the temporal mixing.
+        switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
+            If `True`, switch the spatial and temporal mixing.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        temb_channels: int = 512,
+        eps: float = 1e-6,
+        temporal_eps: Optional[float] = None,
+        merge_factor: float = 0.5,
+        merge_strategy="learned_with_images",
+        switch_spatial_to_temporal_mix: bool = False,
+    ):
+        super().__init__()
+
+        self.spatial_res_block = ResnetBlock2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            eps=eps,
+        )
+
+        self.temporal_res_block = TemporalResnetBlock(
+            in_channels=out_channels if out_channels is not None else in_channels,
+            out_channels=out_channels if out_channels is not None else in_channels,
+            temb_channels=temb_channels,
+            eps=temporal_eps if temporal_eps is not None else eps,
+        )
+
+        self.time_mixer = AlphaBlender(
+            alpha=merge_factor,
+            merge_strategy=merge_strategy,
+            switch_spatial_to_temporal_mix=switch_spatial_to_temporal_mix,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        image_only_indicator: Optional[torch.Tensor] = None,
+    ):
+        num_frames = image_only_indicator.shape[-1]
+        hidden_states = self.spatial_res_block(hidden_states, temb)
+
+        batch_frames, channels, height, width = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states_mix = (
+            hidden_states[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+        )
+        hidden_states = (
+            hidden_states[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+        )
+
+        if temb is not None:
+            temb = temb.reshape(batch_size, num_frames, -1)
+
+        hidden_states = self.temporal_res_block(hidden_states, temb)
+        hidden_states = self.time_mixer(
+            x_spatial=hidden_states_mix,
+            x_temporal=hidden_states,
+            image_only_indicator=image_only_indicator,
+        )
+
+        hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(batch_frames, channels, height, width)
+        return hidden_states
+
+
+class AlphaBlender(nn.Module):
+    r"""
+    A module to blend spatial and temporal features.
+
+    Parameters:
+        alpha (`float`): The initial value of the blending factor.
+        merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
+            The merge strategy to use for the temporal mixing.
+        switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
+            If `True`, switch the spatial and temporal mixing.
+    """
+
+    strategies = ["learned", "fixed", "learned_with_images"]
+
+    def __init__(
+        self,
+        alpha: float,
+        merge_strategy: str = "learned_with_images",
+        switch_spatial_to_temporal_mix: bool = False,
+    ):
+        super().__init__()
+        self.merge_strategy = merge_strategy
+        self.switch_spatial_to_temporal_mix = switch_spatial_to_temporal_mix  # For TemporalVAE
+
+        if merge_strategy not in self.strategies:
+            raise ValueError(f"merge_strategy needs to be in {self.strategies}")
+
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned" or self.merge_strategy == "learned_with_images":
+            self.register_parameter("mix_factor", torch.nn.Parameter(torch.Tensor([alpha])))
+        else:
+            raise ValueError(f"Unknown merge strategy {self.merge_strategy}")
+
+    def get_alpha(self, image_only_indicator: torch.Tensor, ndims: int) -> torch.Tensor:
+        if self.merge_strategy == "fixed":
+            alpha = self.mix_factor
+
+        elif self.merge_strategy == "learned":
+            alpha = torch.sigmoid(self.mix_factor)
+
+        elif self.merge_strategy == "learned_with_images":
+            if image_only_indicator is None:
+                raise ValueError("Please provide image_only_indicator to use learned_with_images merge strategy")
+
+            alpha = torch.where(
+                image_only_indicator.bool(),
+                torch.ones(1, 1, device=image_only_indicator.device),
+                torch.sigmoid(self.mix_factor)[..., None],
+            )
+
+            # (batch, channel, frames, height, width)
+            if ndims == 5:
+                alpha = alpha[:, None, :, None, None]
+            # (batch*frames, height*width, channels)
+            elif ndims == 3:
+                alpha = alpha.reshape(-1)[:, None, None]
+            else:
+                raise ValueError(f"Unexpected ndims {ndims}. Dimensions should be 3 or 5")
+
+        else:
+            raise NotImplementedError
+
+        return alpha
+
+    def forward(
+        self,
+        x_spatial: torch.Tensor,
+        x_temporal: torch.Tensor,
+        image_only_indicator: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        alpha = self.get_alpha(image_only_indicator, x_spatial.ndim)
+        alpha = alpha.to(x_spatial.dtype)
+
+        if self.switch_spatial_to_temporal_mix:
+            alpha = 1.0 - alpha
+
+        x = alpha * x_spatial + (1.0 - alpha) * x_temporal
+        return x

foleycrafter/models/auffusion/transformer_2d.py

@@ -0,0 +1,460 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from .attention import BasicTransformerBlock
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        if self.is_input_continuous:
+            self.in_channels = in_channels
+
+            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+            if use_linear_projection:
+                self.proj_in = linear_cls(in_channels, inner_dim)
+            else:
+                self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
+
+            self.height = sample_size
+            self.width = sample_size
+            self.num_vector_embeds = num_vector_embeds
+            self.num_latent_pixels = self.height * self.width
+
+            self.latent_image_embedding = ImagePositionalEmbeddings(
+                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
+            )
+        elif self.is_input_patches:
+            assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+            self.height = sample_size
+            self.width = sample_size
+
+            self.patch_size = patch_size
+            interpolation_scale = self.config.sample_size // 64  # => 64 (= 512 pixart) has interpolation scale 1
+            interpolation_scale = max(interpolation_scale, 1)
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+                interpolation_scale=interpolation_scale,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                # NOTE: remember to change
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        if self.is_input_continuous:
+            # TODO: should use out_channels for continuous projections
+            if use_linear_projection:
+                self.proj_out = linear_cls(inner_dim, in_channels)
+            else:
+                self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            self.norm_out = nn.LayerNorm(inner_dim)
+            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
+        elif self.is_input_patches and norm_type != "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
+            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+        elif self.is_input_patches and norm_type == "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+            self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 1. Input
+        if self.is_input_continuous:
+            batch, _, height, width = hidden_states.shape
+            inner_dim = hidden_states.shape[1]
+            residual = hidden_states
+
+            hidden_states = self.norm(hidden_states)
+            if not self.use_linear_projection:
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+            else:
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+            self.height, self.width = height, width
+            hidden_states = self.pos_embed(hidden_states)
+
+            if self.adaln_single is not None:
+                if self.use_additional_conditions and added_cond_kwargs is None:
+                    raise ValueError(
+                        "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+                    )
+                batch_size = hidden_states.shape[0]
+                timestep, embedded_timestep = self.adaln_single(
+                    timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+                )
+
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+            output = hidden_states + residual
+        elif self.is_input_vectorized:
+            hidden_states = self.norm_out(hidden_states)
+            logits = self.out(hidden_states)
+            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+            logits = logits.permute(0, 2, 1)
+
+            # log(p(x_0))
+            output = F.log_softmax(logits.double(), dim=1).float()
+
+        if self.is_input_patches:
+            if self.config.norm_type != "ada_norm_single":
+                conditioning = self.transformer_blocks[0].norm1.emb(
+                    timestep, class_labels, hidden_dtype=hidden_states.dtype
+                )
+                shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+                hidden_states = self.proj_out_2(hidden_states)
+            elif self.config.norm_type == "ada_norm_single":
+                shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states)
+                # Modulation
+                hidden_states = hidden_states * (1 + scale) + shift
+                hidden_states = self.proj_out(hidden_states)
+                hidden_states = hidden_states.squeeze(1)
+
+            # unpatchify
+            if self.adaln_single is None:
+                height = width = int(hidden_states.shape[1] ** 0.5)
+            hidden_states = hidden_states.reshape(
+                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+            )
+            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+            output = hidden_states.reshape(
+                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+            )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

foleycrafter/models/onset/__init__.py

@@ -0,0 +1,5 @@
+from .r2plus1d_18 import r2plus1d18KeepTemp
+from .video_onset_net import VideoOnsetNet
+
+
+__all__ = ["r2plus1d18KeepTemp", "VideoOnsetNet"]

foleycrafter/models/onset/r2plus1d_18.py

@@ -0,0 +1,48 @@
+# Copied from specvqgan/onset_baseline/models/r2plus1d_18.py
+
+import torch
+import torch.nn as nn
+
+from .resnet import r2plus1d_18
+
+
+class r2plus1d18KeepTemp(nn.Module):
+    def __init__(self, pretrained=True):
+        super().__init__()
+
+        self.model = r2plus1d_18(pretrained=pretrained)
+
+        self.model.layer2[0].conv1[0][3] = nn.Conv3d(
+            230, 128, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=(1, 0, 0), bias=False
+        )
+        self.model.layer2[0].downsample = nn.Sequential(
+            nn.Conv3d(64, 128, kernel_size=(1, 1, 1), stride=(1, 2, 2), bias=False),
+            nn.BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True),
+        )
+        self.model.layer3[0].conv1[0][3] = nn.Conv3d(
+            460, 256, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=(1, 0, 0), bias=False
+        )
+        self.model.layer3[0].downsample = nn.Sequential(
+            nn.Conv3d(128, 256, kernel_size=(1, 1, 1), stride=(1, 2, 2), bias=False),
+            nn.BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True),
+        )
+        self.model.layer4[0].conv1[0][3] = nn.Conv3d(
+            921, 512, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=(1, 0, 0), bias=False
+        )
+        self.model.layer4[0].downsample = nn.Sequential(
+            nn.Conv3d(256, 512, kernel_size=(1, 1, 1), stride=(1, 2, 2), bias=False),
+            nn.BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True),
+        )
+        self.model.avgpool = nn.AdaptiveAvgPool3d((None, 1, 1))
+        self.model.fc = nn.Identity()
+
+    def forward(self, x):
+        # import pdb; pdb.set_trace()
+        x = self.model(x)
+        return x
+
+
+if __name__ == "__main__":
+    model = r2plus1d18KeepTemp(False).cuda()
+    rand_input = torch.randn((1, 3, 30, 112, 112)).cuda()
+    out = model(rand_input)

foleycrafter/models/onset/resnet.py

@@ -0,0 +1,323 @@
+# Copied from specvqgan/onset_baseline/models/resnet.py
+import torch.nn as nn
+from torch.hub import load_state_dict_from_url
+
+
+__all__ = ["r3d_18", "mc3_18", "r2plus1d_18"]
+
+model_urls = {
+    "r3d_18": "https://download.pytorch.org/models/r3d_18-b3b3357e.pth",
+    "mc3_18": "https://download.pytorch.org/models/mc3_18-a90a0ba3.pth",
+    "r2plus1d_18": "https://download.pytorch.org/models/r2plus1d_18-91a641e6.pth",
+}
+
+
+class Conv3DSimple(nn.Conv3d):
+    def __init__(self, in_planes, out_planes, midplanes=None, stride=1, padding=1):
+        super(Conv3DSimple, self).__init__(
+            in_channels=in_planes,
+            out_channels=out_planes,
+            kernel_size=(3, 3, 3),
+            stride=stride,
+            padding=padding,
+            bias=False,
+        )
+
+    @staticmethod
+    def get_downsample_stride(stride):
+        return stride, stride, stride
+
+
+class Conv2Plus1D(nn.Sequential):
+    def __init__(self, in_planes, out_planes, midplanes, stride=1, padding=1):
+        super(Conv2Plus1D, self).__init__(
+            nn.Conv3d(
+                in_planes,
+                midplanes,
+                kernel_size=(1, 3, 3),
+                stride=(1, stride, stride),
+                padding=(0, padding, padding),
+                bias=False,
+            ),
+            nn.BatchNorm3d(midplanes),
+            nn.ReLU(inplace=True),
+            nn.Conv3d(
+                midplanes,
+                out_planes,
+                kernel_size=(3, 1, 1),
+                stride=(stride, 1, 1),
+                padding=(padding, 0, 0),
+                bias=False,
+            ),
+        )
+
+    @staticmethod
+    def get_downsample_stride(stride):
+        return stride, stride, stride
+
+
+class Conv3DNoTemporal(nn.Conv3d):
+    def __init__(self, in_planes, out_planes, midplanes=None, stride=1, padding=1):
+        super(Conv3DNoTemporal, self).__init__(
+            in_channels=in_planes,
+            out_channels=out_planes,
+            kernel_size=(1, 3, 3),
+            stride=(1, stride, stride),
+            padding=(0, padding, padding),
+            bias=False,
+        )
+
+    @staticmethod
+    def get_downsample_stride(stride):
+        return 1, stride, stride
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, conv_builder, stride=1, downsample=None):
+        midplanes = (inplanes * planes * 3 * 3 * 3) // (inplanes * 3 * 3 + 3 * planes)
+
+        super(BasicBlock, self).__init__()
+        self.conv1 = nn.Sequential(
+            conv_builder(inplanes, planes, midplanes, stride), nn.BatchNorm3d(planes), nn.ReLU(inplace=True)
+        )
+        self.conv2 = nn.Sequential(conv_builder(planes, planes, midplanes), nn.BatchNorm3d(planes))
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.conv2(out)
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, conv_builder, stride=1, downsample=None):
+        super(Bottleneck, self).__init__()
+        midplanes = (inplanes * planes * 3 * 3 * 3) // (inplanes * 3 * 3 + 3 * planes)
+
+        # 1x1x1
+        self.conv1 = nn.Sequential(
+            nn.Conv3d(inplanes, planes, kernel_size=1, bias=False), nn.BatchNorm3d(planes), nn.ReLU(inplace=True)
+        )
+        # Second kernel
+        self.conv2 = nn.Sequential(
+            conv_builder(planes, planes, midplanes, stride), nn.BatchNorm3d(planes), nn.ReLU(inplace=True)
+        )
+
+        # 1x1x1
+        self.conv3 = nn.Sequential(
+            nn.Conv3d(planes, planes * self.expansion, kernel_size=1, bias=False),
+            nn.BatchNorm3d(planes * self.expansion),
+        )
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.conv2(out)
+        out = self.conv3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class BasicStem(nn.Sequential):
+    """The default conv-batchnorm-relu stem"""
+
+    def __init__(self):
+        super(BasicStem, self).__init__(
+            nn.Conv3d(3, 64, kernel_size=(3, 7, 7), stride=(1, 2, 2), padding=(1, 3, 3), bias=False),
+            nn.BatchNorm3d(64),
+            nn.ReLU(inplace=True),
+        )
+
+
+class R2Plus1dStem(nn.Sequential):
+    """R(2+1)D stem is different than the default one as it uses separated 3D convolution"""
+
+    def __init__(self):
+        super(R2Plus1dStem, self).__init__(
+            nn.Conv3d(3, 45, kernel_size=(1, 7, 7), stride=(1, 2, 2), padding=(0, 3, 3), bias=False),
+            nn.BatchNorm3d(45),
+            nn.ReLU(inplace=True),
+            nn.Conv3d(45, 64, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=(1, 0, 0), bias=False),
+            nn.BatchNorm3d(64),
+            nn.ReLU(inplace=True),
+        )
+
+
+class VideoResNet(nn.Module):
+    def __init__(self, block, conv_makers, layers, stem, num_classes=400, zero_init_residual=False):
+        """Generic resnet video generator.
+        Args:
+            block (nn.Module): resnet building block
+            conv_makers (list(functions)): generator function for each layer
+            layers (List[int]): number of blocks per layer
+            stem (nn.Module, optional): Resnet stem, if None, defaults to conv-bn-relu. Defaults to None.
+            num_classes (int, optional): Dimension of the final FC layer. Defaults to 400.
+            zero_init_residual (bool, optional): Zero init bottleneck residual BN. Defaults to False.
+        """
+        super(VideoResNet, self).__init__()
+        self.inplanes = 64
+
+        self.stem = stem()
+
+        self.layer1 = self._make_layer(block, conv_makers[0], 64, layers[0], stride=1)
+        self.layer2 = self._make_layer(block, conv_makers[1], 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, conv_makers[2], 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, conv_makers[3], 512, layers[3], stride=2)
+
+        self.avgpool = nn.AdaptiveAvgPool3d((1, 1, 1))
+        self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        # init weights
+        self._initialize_weights()
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)
+
+    def forward(self, x):
+        x = self.stem(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        # Flatten the layer to fc
+        # x = x.flatten(1)
+        # x = self.fc(x)
+        N = x.shape[0]
+        x = x.squeeze()
+        if N == 1:
+            x = x[None]
+
+        return x
+
+    def _make_layer(self, block, conv_builder, planes, blocks, stride=1):
+        downsample = None
+
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            ds_stride = conv_builder.get_downsample_stride(stride)
+            downsample = nn.Sequential(
+                nn.Conv3d(self.inplanes, planes * block.expansion, kernel_size=1, stride=ds_stride, bias=False),
+                nn.BatchNorm3d(planes * block.expansion),
+            )
+        layers = []
+        layers.append(block(self.inplanes, planes, conv_builder, stride, downsample))
+
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes, conv_builder))
+
+        return nn.Sequential(*layers)
+
+    def _initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv3d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.BatchNorm3d):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                nn.init.constant_(m.bias, 0)
+
+
+def _video_resnet(arch, pretrained=False, progress=True, **kwargs):
+    model = VideoResNet(**kwargs)
+
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+
+def r3d_18(pretrained=False, progress=True, **kwargs):
+    """Construct 18 layer Resnet3D model as in
+    https://arxiv.org/abs/1711.11248
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on Kinetics-400
+        progress (bool): If True, displays a progress bar of the download to stderr
+    Returns:
+        nn.Module: R3D-18 network
+    """
+
+    return _video_resnet(
+        "r3d_18",
+        pretrained,
+        progress,
+        block=BasicBlock,
+        conv_makers=[Conv3DSimple] * 4,
+        layers=[2, 2, 2, 2],
+        stem=BasicStem,
+        **kwargs,
+    )
+
+
+def mc3_18(pretrained=False, progress=True, **kwargs):
+    """Constructor for 18 layer Mixed Convolution network as in
+    https://arxiv.org/abs/1711.11248
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on Kinetics-400
+        progress (bool): If True, displays a progress bar of the download to stderr
+    Returns:
+        nn.Module: MC3 Network definition
+    """
+    return _video_resnet(
+        "mc3_18",
+        pretrained,
+        progress,
+        block=BasicBlock,
+        conv_makers=[Conv3DSimple] + [Conv3DNoTemporal] * 3,
+        layers=[2, 2, 2, 2],
+        stem=BasicStem,
+        **kwargs,
+    )
+
+
+def r2plus1d_18(pretrained=False, progress=True, **kwargs):
+    """Constructor for the 18 layer deep R(2+1)D network as in
+    https://arxiv.org/abs/1711.11248
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on Kinetics-400
+        progress (bool): If True, displays a progress bar of the download to stderr
+    Returns:
+        nn.Module: R(2+1)D-18 network
+    """
+    return _video_resnet(
+        "r2plus1d_18",
+        pretrained,
+        progress,
+        block=BasicBlock,
+        conv_makers=[Conv2Plus1D] * 4,
+        layers=[2, 2, 2, 2],
+        stem=R2Plus1dStem,
+        **kwargs,
+    )

foleycrafter/models/onset/torch_utils.py

@@ -0,0 +1,113 @@
+# Copied from https://github.com/XYPB/CondFoleyGen/blob/main/specvqgan/onset_baseline/utils/torch_utils.py
+import os
+import sys
+from collections import OrderedDict
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader
+
+from ... import data
+
+
+# ---------------------------------------------------- #
+def load_model(cp_path, net, device=None, strict=True):
+    if not device:
+        device = torch.device("cpu")
+    if os.path.isfile(cp_path):
+        print("=> loading checkpoint '{}'".format(cp_path))
+        checkpoint = torch.load(cp_path, map_location=device)
+
+        # check if there is module
+        if list(checkpoint["state_dict"].keys())[0][:7] == "module.":
+            state_dict = OrderedDict()
+            for k, v in checkpoint["state_dict"].items():
+                name = k[7:]
+                state_dict[name] = v
+        else:
+            state_dict = checkpoint["state_dict"]
+        net.load_state_dict(state_dict, strict=strict)
+
+        print("=> loaded checkpoint '{}' (epoch {})".format(cp_path, checkpoint["epoch"]))
+        start_epoch = checkpoint["epoch"]
+    else:
+        print("=> no checkpoint found at '{}'".format(cp_path))
+        start_epoch = 0
+        sys.exit()
+
+    return net, start_epoch
+
+
+# ---------------------------------------------------- #
+def binary_acc(pred, target, threshold):
+    pred = pred > threshold
+    acc = np.sum(pred == target) / target.shape[0]
+    return acc
+
+
+def calc_acc(prob, labels, k):
+    pred = torch.argsort(prob, dim=-1, descending=True)[..., :k]
+    top_k_acc = torch.sum(pred == labels.view(-1, 1)).float() / labels.size(0)
+    return top_k_acc
+
+
+# ---------------------------------------------------- #
+
+
+def get_dataloader(args, pr, split="train", shuffle=False, drop_last=False, batch_size=None):
+    data_loader = getattr(data, pr.dataloader)
+    if split == "train":
+        read_list = pr.list_train
+    elif split == "val":
+        read_list = pr.list_val
+    elif split == "test":
+        read_list = pr.list_test
+    dataset = data_loader(args, pr, read_list, split=split)
+    batch_size = batch_size if batch_size else args.batch_size
+    dataset.getitem_test(1)
+    loader = DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=shuffle,
+        num_workers=args.num_workers,
+        pin_memory=True,
+        drop_last=drop_last,
+    )
+
+    return dataset, loader
+
+
+# ---------------------------------------------------- #
+def make_optimizer(model, args):
+    """
+    Args:
+        model: NN to train
+    Returns:
+        optimizer: pytorch optmizer for updating the given model parameters.
+    """
+    if args.optim == "SGD":
+        optimizer = torch.optim.SGD(
+            filter(lambda p: p.requires_grad, model.parameters()),
+            lr=args.lr,
+            momentum=args.momentum,
+            weight_decay=args.weight_decay,
+            nesterov=False,
+        )
+    elif args.optim == "Adam":
+        optimizer = torch.optim.Adam(
+            filter(lambda p: p.requires_grad, model.parameters()),
+            lr=args.lr,
+            weight_decay=args.weight_decay,
+        )
+    return optimizer
+
+
+def adjust_learning_rate(optimizer, epoch, args):
+    """Decay the learning rate based on schedule"""
+    lr = args.lr
+    if args.schedule == "cos":  # cosine lr schedule
+        lr *= 0.5 * (1.0 + np.cos(np.pi * epoch / args.epochs))
+    elif args.schedule == "none":  # no lr schedule
+        lr = args.lr
+    for param_group in optimizer.param_groups:
+        param_group["lr"] = lr

foleycrafter/models/onset/video_onset_net.py

@@ -0,0 +1,31 @@
+# Copied from specvqgan/onset_baseline/models/video_onset_net.py
+
+import torch
+import torch.nn as nn
+
+from .r2plus1d_18 import r2plus1d18KeepTemp
+
+
+class VideoOnsetNet(nn.Module):
+    # Video Onset detection network
+    def __init__(self, pretrained):
+        super(VideoOnsetNet, self).__init__()
+        self.net = r2plus1d18KeepTemp(pretrained=pretrained)
+        self.fc = nn.Sequential(nn.Linear(512, 128), nn.ReLU(True), nn.Linear(128, 1))
+
+    def forward(self, inputs, loss=False, evaluate=False):
+        # import pdb; pdb.set_trace()
+        x = inputs["frames"]
+        x = self.net(x)
+        x = x.transpose(-1, -2)
+        x = self.fc(x)
+        x = x.squeeze(-1)
+
+        return x
+
+
+if __name__ == "__main__":
+    model = VideoOnsetNet(False).cuda()
+    rand_input = torch.randn((1, 3, 30, 112, 112)).cuda()
+    inputs = {"frames": rand_input}
+    out = model(inputs)

foleycrafter/models/time_detector/model.py

@@ -0,0 +1,17 @@
+import torch.nn as nn
+
+from ..onset import VideoOnsetNet
+
+
+class TimeDetector(nn.Module):
+    def __init__(self, video_length=150, audio_length=1024):
+        super(TimeDetector, self).__init__()
+        self.pred_net = VideoOnsetNet(pretrained=False)
+        self.soft_fn = nn.Tanh()
+        self.up_sampler = nn.Linear(video_length, audio_length)
+
+    def forward(self, inputs):
+        x = self.pred_net(inputs)
+        x = self.up_sampler(x)
+        x = self.soft_fn(x)
+        return x

foleycrafter/models/time_detector/resnet.py

@@ -0,0 +1,322 @@
+import torch.nn as nn
+from torch.hub import load_state_dict_from_url
+
+
+__all__ = ["r3d_18", "mc3_18", "r2plus1d_18"]
+
+model_urls = {
+    "r3d_18": "https://download.pytorch.org/models/r3d_18-b3b3357e.pth",
+    "mc3_18": "https://download.pytorch.org/models/mc3_18-a90a0ba3.pth",
+    "r2plus1d_18": "https://download.pytorch.org/models/r2plus1d_18-91a641e6.pth",
+}
+
+
+class Conv3DSimple(nn.Conv3d):
+    def __init__(self, in_planes, out_planes, midplanes=None, stride=1, padding=1):
+        super(Conv3DSimple, self).__init__(
+            in_channels=in_planes,
+            out_channels=out_planes,
+            kernel_size=(3, 3, 3),
+            stride=stride,
+            padding=padding,
+            bias=False,
+        )
+
+    @staticmethod
+    def get_downsample_stride(stride):
+        return stride, stride, stride
+
+
+class Conv2Plus1D(nn.Sequential):
+    def __init__(self, in_planes, out_planes, midplanes, stride=1, padding=1):
+        super(Conv2Plus1D, self).__init__(
+            nn.Conv3d(
+                in_planes,
+                midplanes,
+                kernel_size=(1, 3, 3),
+                stride=(1, stride, stride),
+                padding=(0, padding, padding),
+                bias=False,
+            ),
+            nn.BatchNorm3d(midplanes),
+            nn.ReLU(inplace=True),
+            nn.Conv3d(
+                midplanes,
+                out_planes,
+                kernel_size=(3, 1, 1),
+                stride=(stride, 1, 1),
+                padding=(padding, 0, 0),
+                bias=False,
+            ),
+        )
+
+    @staticmethod
+    def get_downsample_stride(stride):
+        return stride, stride, stride
+
+
+class Conv3DNoTemporal(nn.Conv3d):
+    def __init__(self, in_planes, out_planes, midplanes=None, stride=1, padding=1):
+        super(Conv3DNoTemporal, self).__init__(
+            in_channels=in_planes,
+            out_channels=out_planes,
+            kernel_size=(1, 3, 3),
+            stride=(1, stride, stride),
+            padding=(0, padding, padding),
+            bias=False,
+        )
+
+    @staticmethod
+    def get_downsample_stride(stride):
+        return 1, stride, stride
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, conv_builder, stride=1, downsample=None):
+        midplanes = (inplanes * planes * 3 * 3 * 3) // (inplanes * 3 * 3 + 3 * planes)
+
+        super(BasicBlock, self).__init__()
+        self.conv1 = nn.Sequential(
+            conv_builder(inplanes, planes, midplanes, stride), nn.BatchNorm3d(planes), nn.ReLU(inplace=True)
+        )
+        self.conv2 = nn.Sequential(conv_builder(planes, planes, midplanes), nn.BatchNorm3d(planes))
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.conv2(out)
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, conv_builder, stride=1, downsample=None):
+        super(Bottleneck, self).__init__()
+        midplanes = (inplanes * planes * 3 * 3 * 3) // (inplanes * 3 * 3 + 3 * planes)
+
+        # 1x1x1
+        self.conv1 = nn.Sequential(
+            nn.Conv3d(inplanes, planes, kernel_size=1, bias=False), nn.BatchNorm3d(planes), nn.ReLU(inplace=True)
+        )
+        # Second kernel
+        self.conv2 = nn.Sequential(
+            conv_builder(planes, planes, midplanes, stride), nn.BatchNorm3d(planes), nn.ReLU(inplace=True)
+        )
+
+        # 1x1x1
+        self.conv3 = nn.Sequential(
+            nn.Conv3d(planes, planes * self.expansion, kernel_size=1, bias=False),
+            nn.BatchNorm3d(planes * self.expansion),
+        )
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.conv2(out)
+        out = self.conv3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class BasicStem(nn.Sequential):
+    """The default conv-batchnorm-relu stem"""
+
+    def __init__(self):
+        super(BasicStem, self).__init__(
+            nn.Conv3d(3, 64, kernel_size=(3, 7, 7), stride=(1, 2, 2), padding=(1, 3, 3), bias=False),
+            nn.BatchNorm3d(64),
+            nn.ReLU(inplace=True),
+        )
+
+
+class R2Plus1dStem(nn.Sequential):
+    """R(2+1)D stem is different than the default one as it uses separated 3D convolution"""
+
+    def __init__(self):
+        super(R2Plus1dStem, self).__init__(
+            nn.Conv3d(3, 45, kernel_size=(1, 7, 7), stride=(1, 2, 2), padding=(0, 3, 3), bias=False),
+            nn.BatchNorm3d(45),
+            nn.ReLU(inplace=True),
+            nn.Conv3d(45, 64, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=(1, 0, 0), bias=False),
+            nn.BatchNorm3d(64),
+            nn.ReLU(inplace=True),
+        )
+
+
+class VideoResNet(nn.Module):
+    def __init__(self, block, conv_makers, layers, stem, num_classes=400, zero_init_residual=False):
+        """Generic resnet video generator.
+        Args:
+            block (nn.Module): resnet building block
+            conv_makers (list(functions)): generator function for each layer
+            layers (List[int]): number of blocks per layer
+            stem (nn.Module, optional): Resnet stem, if None, defaults to conv-bn-relu. Defaults to None.
+            num_classes (int, optional): Dimension of the final FC layer. Defaults to 400.
+            zero_init_residual (bool, optional): Zero init bottleneck residual BN. Defaults to False.
+        """
+        super(VideoResNet, self).__init__()
+        self.inplanes = 64
+
+        self.stem = stem()
+
+        self.layer1 = self._make_layer(block, conv_makers[0], 64, layers[0], stride=1)
+        self.layer2 = self._make_layer(block, conv_makers[1], 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, conv_makers[2], 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, conv_makers[3], 512, layers[3], stride=2)
+
+        self.avgpool = nn.AdaptiveAvgPool3d((1, 1, 1))
+        self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        # init weights
+        self._initialize_weights()
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)
+
+    def forward(self, x):
+        x = self.stem(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        # Flatten the layer to fc
+        # x = x.flatten(1)
+        # x = self.fc(x)
+        N = x.shape[0]
+        x = x.squeeze()
+        if N == 1:
+            x = x[None]
+
+        return x
+
+    def _make_layer(self, block, conv_builder, planes, blocks, stride=1):
+        downsample = None
+
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            ds_stride = conv_builder.get_downsample_stride(stride)
+            downsample = nn.Sequential(
+                nn.Conv3d(self.inplanes, planes * block.expansion, kernel_size=1, stride=ds_stride, bias=False),
+                nn.BatchNorm3d(planes * block.expansion),
+            )
+        layers = []
+        layers.append(block(self.inplanes, planes, conv_builder, stride, downsample))
+
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes, conv_builder))
+
+        return nn.Sequential(*layers)
+
+    def _initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv3d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.BatchNorm3d):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                nn.init.constant_(m.bias, 0)
+
+
+def _video_resnet(arch, pretrained=False, progress=True, **kwargs):
+    model = VideoResNet(**kwargs)
+
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+
+def r3d_18(pretrained=False, progress=True, **kwargs):
+    """Construct 18 layer Resnet3D model as in
+    https://arxiv.org/abs/1711.11248
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on Kinetics-400
+        progress (bool): If True, displays a progress bar of the download to stderr
+    Returns:
+        nn.Module: R3D-18 network
+    """
+
+    return _video_resnet(
+        "r3d_18",
+        pretrained,
+        progress,
+        block=BasicBlock,
+        conv_makers=[Conv3DSimple] * 4,
+        layers=[2, 2, 2, 2],
+        stem=BasicStem,
+        **kwargs,
+    )
+
+
+def mc3_18(pretrained=False, progress=True, **kwargs):
+    """Constructor for 18 layer Mixed Convolution network as in
+    https://arxiv.org/abs/1711.11248
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on Kinetics-400
+        progress (bool): If True, displays a progress bar of the download to stderr
+    Returns:
+        nn.Module: MC3 Network definition
+    """
+    return _video_resnet(
+        "mc3_18",
+        pretrained,
+        progress,
+        block=BasicBlock,
+        conv_makers=[Conv3DSimple] + [Conv3DNoTemporal] * 3,
+        layers=[2, 2, 2, 2],
+        stem=BasicStem,
+        **kwargs,
+    )
+
+
+def r2plus1d_18(pretrained=False, progress=True, **kwargs):
+    """Constructor for the 18 layer deep R(2+1)D network as in
+    https://arxiv.org/abs/1711.11248
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on Kinetics-400
+        progress (bool): If True, displays a progress bar of the download to stderr
+    Returns:
+        nn.Module: R(2+1)D-18 network
+    """
+    return _video_resnet(
+        "r2plus1d_18",
+        pretrained,
+        progress,
+        block=BasicBlock,
+        conv_makers=[Conv2Plus1D] * 4,
+        layers=[2, 2, 2, 2],
+        stem=R2Plus1dStem,
+        **kwargs,
+    )

foleycrafter/utils/converter.py

@@ -0,0 +1,494 @@
+# Copy from https://github.com/happylittlecat2333/Auffusion/blob/main/converter.py
+import json
+import os
+import random
+
+import librosa
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+
+# from librosa.util import normalize
+from scipy.io.wavfile import read
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+
+MAX_WAV_VALUE = 32768.0
+
+
+def load_wav(full_path):
+    sampling_rate, data = read(full_path)
+    return data, sampling_rate
+
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def mel_spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
+    if torch.min(y) < -1.0:
+        print("min value is ", torch.min(y))
+    if torch.max(y) > 1.0:
+        print("max value is ", torch.max(y))
+
+    global mel_basis, hann_window
+    if fmax not in mel_basis:
+        mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
+        mel_basis[str(fmax) + "_" + str(y.device)] = torch.from_numpy(mel).float().to(y.device)
+        hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
+
+    y = torch.nn.functional.pad(
+        y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect"
+    )
+    y = y.squeeze(1)
+
+    # complex tensor as default, then use view_as_real for future pytorch compatibility
+    spec = torch.stft(
+        y,
+        n_fft,
+        hop_length=hop_size,
+        win_length=win_size,
+        window=hann_window[str(y.device)],
+        center=center,
+        pad_mode="reflect",
+        normalized=False,
+        onesided=True,
+        return_complex=True,
+    )
+    spec = torch.view_as_real(spec)
+    spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+
+    spec = torch.matmul(mel_basis[str(fmax) + "_" + str(y.device)], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec
+
+
+def spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
+    if torch.min(y) < -1.0:
+        print("min value is ", torch.min(y))
+    if torch.max(y) > 1.0:
+        print("max value is ", torch.max(y))
+
+    global hann_window
+    hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
+
+    y = torch.nn.functional.pad(
+        y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect"
+    )
+    y = y.squeeze(1)
+
+    # complex tensor as default, then use view_as_real for future pytorch compatibility
+    spec = torch.stft(
+        y,
+        n_fft,
+        hop_length=hop_size,
+        win_length=win_size,
+        window=hann_window[str(y.device)],
+        center=center,
+        pad_mode="reflect",
+        normalized=False,
+        onesided=True,
+        return_complex=True,
+    )
+    spec = torch.view_as_real(spec)
+    spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+
+    return spec
+
+
+def normalize_spectrogram(
+    spectrogram: torch.Tensor,
+    max_value: float = 200,
+    min_value: float = 1e-5,
+    power: float = 1.0,
+    inverse: bool = False,
+) -> torch.Tensor:
+    # Rescale to 0-1
+    max_value = np.log(max_value)  # 5.298317366548036
+    min_value = np.log(min_value)  # -11.512925464970229
+
+    assert spectrogram.max() <= max_value and spectrogram.min() >= min_value
+
+    data = (spectrogram - min_value) / (max_value - min_value)
+
+    # Invert
+    if inverse:
+        data = 1 - data
+
+    # Apply the power curve
+    data = torch.pow(data, power)
+
+    # 1D -> 3D
+    data = data.repeat(3, 1, 1)
+
+    # Flip Y axis: image origin at the top-left corner, spectrogram origin at the bottom-left corner
+    data = torch.flip(data, [1])
+
+    return data
+
+
+def denormalize_spectrogram(
+    data: torch.Tensor,
+    max_value: float = 200,
+    min_value: float = 1e-5,
+    power: float = 1,
+    inverse: bool = False,
+) -> torch.Tensor:
+    max_value = np.log(max_value)
+    min_value = np.log(min_value)
+
+    # Flip Y axis: image origin at the top-left corner, spectrogram origin at the bottom-left corner
+    data = torch.flip(data, [1])
+
+    assert len(data.shape) == 3, "Expected 3 dimensions, got {}".format(len(data.shape))
+
+    if data.shape[0] == 1:
+        data = data.repeat(3, 1, 1)
+
+    assert data.shape[0] == 3, "Expected 3 channels, got {}".format(data.shape[0])
+    data = data[0]
+
+    # Reverse the power curve
+    data = torch.pow(data, 1 / power)
+
+    # Invert
+    if inverse:
+        data = 1 - data
+
+    # Rescale to max value
+    spectrogram = data * (max_value - min_value) + min_value
+
+    return spectrogram
+
+
+def get_mel_spectrogram_from_audio(audio, device="cpu"):
+    audio = audio / MAX_WAV_VALUE
+    audio = librosa.util.normalize(audio) * 0.95
+    # print(' >>> normalize done <<< ')
+
+    audio = torch.FloatTensor(audio)
+    audio = audio.unsqueeze(0)
+
+    waveform = audio.to(device)
+    spec = mel_spectrogram(
+        waveform,
+        n_fft=2048,
+        num_mels=256,
+        sampling_rate=16000,
+        hop_size=160,
+        win_size=1024,
+        fmin=0,
+        fmax=8000,
+        center=False,
+    )
+    return audio, spec
+
+
+LRELU_SLOPE = 0.1
+MAX_WAV_VALUE = 32768.0
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+def get_config(config_path):
+    config = json.loads(open(config_path).read())
+    config = AttrDict(config)
+    return config
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def apply_weight_norm(m):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        weight_norm(m)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.h = h
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1))
+                ),
+                weight_norm(
+                    Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1))
+                ),
+                weight_norm(
+                    Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1))
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.h = h
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(init_weights)
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Generator(torch.nn.Module):
+    def __init__(self, h):
+        super(Generator, self).__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
+        )  # change: 80 --> 512
+        resblock = ResBlock1 if h.resblock == "1" else ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            if (k - u) % 2 == 0:
+                self.ups.append(
+                    weight_norm(
+                        ConvTranspose1d(
+                            h.upsample_initial_channel // (2**i),
+                            h.upsample_initial_channel // (2 ** (i + 1)),
+                            k,
+                            u,
+                            padding=(k - u) // 2,
+                        )
+                    )
+                )
+            else:
+                self.ups.append(
+                    weight_norm(
+                        ConvTranspose1d(
+                            h.upsample_initial_channel // (2**i),
+                            h.upsample_initial_channel // (2 ** (i + 1)),
+                            k,
+                            u,
+                            padding=(k - u) // 2 + 1,
+                            output_padding=1,
+                        )
+                    )
+                )
+
+            # self.ups.append(weight_norm(
+            #     ConvTranspose1d(h.upsample_initial_channel//(2**i), h.upsample_initial_channel//(2**(i+1)),
+            #                     k, u, padding=(k-u)//2)))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, subfolder=None):
+        if subfolder is not None:
+            pretrained_model_name_or_path = os.path.join(pretrained_model_name_or_path, subfolder)
+        config_path = os.path.join(pretrained_model_name_or_path, "config.json")
+        ckpt_path = os.path.join(pretrained_model_name_or_path, "vocoder.pt")
+
+        config = get_config(config_path)
+        vocoder = cls(config)
+
+        state_dict_g = torch.load(ckpt_path)
+        vocoder.load_state_dict(state_dict_g["generator"])
+        vocoder.eval()
+        vocoder.remove_weight_norm()
+        return vocoder
+
+    @torch.no_grad()
+    def inference(self, mels, lengths=None):
+        self.eval()
+        with torch.no_grad():
+            wavs = self(mels).squeeze(1)
+
+        wavs = (wavs.cpu().numpy() * MAX_WAV_VALUE).astype("int16")
+
+        if lengths is not None:
+            wavs = wavs[:, :lengths]
+
+        return wavs
+
+
+def normalize(images):
+    """
+    Normalize an image array to [-1,1].
+    """
+    if images.min() >= 0:
+        return 2.0 * images - 1.0
+    else:
+        return images
+
+
+def pad_spec(spec, spec_length, pad_value=0, random_crop=True):  # spec: [3, mel_dim, spec_len]
+    assert spec_length % 8 == 0, "spec_length must be divisible by 8"
+    if spec.shape[-1] < spec_length:
+        # pad spec to spec_length
+        spec = F.pad(spec, (0, spec_length - spec.shape[-1]), value=pad_value)
+    else:
+        # random crop
+        if random_crop:
+            start = random.randint(0, spec.shape[-1] - spec_length)
+            spec = spec[:, :, start : start + spec_length]
+        else:
+            spec = spec[:, :, :spec_length]
+    return spec

foleycrafter/utils/spec_to_mel.py

@@ -0,0 +1,399 @@
+import librosa.util as librosa_util
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torchaudio
+from librosa.filters import mel as librosa_mel_fn
+from librosa.util import pad_center, tiny
+from scipy.signal import get_window
+
+
+# spectrogram to mel
+
+
+class STFT(torch.nn.Module):
+    """adapted from Prem Seetharaman's https://github.com/pseeth/pytorch-stft"""
+
+    def __init__(self, filter_length, hop_length, win_length, window="hann"):
+        super(STFT, self).__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.window = window
+        self.forward_transform = None
+        scale = self.filter_length / self.hop_length
+        fourier_basis = np.fft.fft(np.eye(self.filter_length))
+
+        cutoff = int((self.filter_length / 2 + 1))
+        fourier_basis = np.vstack([np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])])
+
+        forward_basis = torch.FloatTensor(fourier_basis[:, None, :])
+        inverse_basis = torch.FloatTensor(np.linalg.pinv(scale * fourier_basis).T[:, None, :])
+
+        if window is not None:
+            assert filter_length >= win_length
+            # get window and zero center pad it to filter_length
+            fft_window = get_window(window, win_length, fftbins=True)
+            fft_window = pad_center(fft_window, filter_length)
+            fft_window = torch.from_numpy(fft_window).float()
+
+            # window the bases
+            forward_basis *= fft_window
+            inverse_basis *= fft_window
+
+        self.register_buffer("forward_basis", forward_basis.float())
+        self.register_buffer("inverse_basis", inverse_basis.float())
+
+    def transform(self, input_data):
+        num_batches = input_data.size(0)
+        num_samples = input_data.size(1)
+
+        self.num_samples = num_samples
+
+        # similar to librosa, reflect-pad the input
+        input_data = input_data.view(num_batches, 1, num_samples)
+        input_data = F.pad(
+            input_data.unsqueeze(1),
+            (int(self.filter_length / 2), int(self.filter_length / 2), 0, 0),
+            mode="reflect",
+        )
+        input_data = input_data.squeeze(1)
+
+        forward_transform = F.conv1d(
+            input_data,
+            torch.autograd.Variable(self.forward_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        ).cpu()
+
+        cutoff = int((self.filter_length / 2) + 1)
+        real_part = forward_transform[:, :cutoff, :]
+        imag_part = forward_transform[:, cutoff:, :]
+
+        magnitude = torch.sqrt(real_part**2 + imag_part**2)
+        phase = torch.autograd.Variable(torch.atan2(imag_part.data, real_part.data))
+
+        return magnitude, phase
+
+    def inverse(self, magnitude, phase):
+        recombine_magnitude_phase = torch.cat([magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1)
+
+        inverse_transform = F.conv_transpose1d(
+            recombine_magnitude_phase,
+            torch.autograd.Variable(self.inverse_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        )
+
+        if self.window is not None:
+            window_sum = window_sumsquare(
+                self.window,
+                magnitude.size(-1),
+                hop_length=self.hop_length,
+                win_length=self.win_length,
+                n_fft=self.filter_length,
+                dtype=np.float32,
+            )
+            # remove modulation effects
+            approx_nonzero_indices = torch.from_numpy(np.where(window_sum > tiny(window_sum))[0])
+            window_sum = torch.autograd.Variable(torch.from_numpy(window_sum), requires_grad=False)
+            window_sum = window_sum
+            inverse_transform[:, :, approx_nonzero_indices] /= window_sum[approx_nonzero_indices]
+
+            # scale by hop ratio
+            inverse_transform *= float(self.filter_length) / self.hop_length
+
+        inverse_transform = inverse_transform[:, :, int(self.filter_length / 2) :]
+        inverse_transform = inverse_transform[:, :, : -int(self.filter_length / 2) :]
+
+        return inverse_transform
+
+    def forward(self, input_data):
+        self.magnitude, self.phase = self.transform(input_data)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+
+
+def window_sumsquare(
+    window,
+    n_frames,
+    hop_length,
+    win_length,
+    n_fft,
+    dtype=np.float32,
+    norm=None,
+):
+    """
+    # from librosa 0.6
+    Compute the sum-square envelope of a window function at a given hop length.
+
+    This is used to estimate modulation effects induced by windowing
+    observations in short-time fourier transforms.
+
+    Parameters
+    ----------
+    window : string, tuple, number, callable, or list-like
+        Window specification, as in `get_window`
+
+    n_frames : int > 0
+        The number of analysis frames
+
+    hop_length : int > 0
+        The number of samples to advance between frames
+
+    win_length : [optional]
+        The length of the window function.  By default, this matches `n_fft`.
+
+    n_fft : int > 0
+        The length of each analysis frame.
+
+    dtype : np.dtype
+        The data type of the output
+
+    Returns
+    -------
+    wss : np.ndarray, shape=`(n_fft + hop_length * (n_frames - 1))`
+        The sum-squared envelope of the window function
+    """
+    if win_length is None:
+        win_length = n_fft
+
+    n = n_fft + hop_length * (n_frames - 1)
+    x = np.zeros(n, dtype=dtype)
+
+    # Compute the squared window at the desired length
+    win_sq = get_window(window, win_length, fftbins=True)
+    win_sq = librosa_util.normalize(win_sq, norm=norm) ** 2
+    win_sq = librosa_util.pad_center(win_sq, n_fft)
+
+    # Fill the envelope
+    for i in range(n_frames):
+        sample = i * hop_length
+        x[sample : min(n, sample + n_fft)] += win_sq[: max(0, min(n_fft, n - sample))]
+    return x
+
+
+def griffin_lim(magnitudes, stft_fn, n_iters=30):
+    """
+    PARAMS
+    ------
+    magnitudes: spectrogram magnitudes
+    stft_fn: STFT class with transform (STFT) and inverse (ISTFT) methods
+    """
+
+    angles = np.angle(np.exp(2j * np.pi * np.random.rand(*magnitudes.size())))
+    angles = angles.astype(np.float32)
+    angles = torch.autograd.Variable(torch.from_numpy(angles))
+    signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+
+    for i in range(n_iters):
+        _, angles = stft_fn.transform(signal)
+        signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+    return signal
+
+
+def dynamic_range_compression(x, normalize_fun=torch.log, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return normalize_fun(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return torch.exp(x) / C
+
+
+class TacotronSTFT(torch.nn.Module):
+    def __init__(
+        self,
+        filter_length,
+        hop_length,
+        win_length,
+        n_mel_channels,
+        sampling_rate,
+        mel_fmin,
+        mel_fmax,
+    ):
+        super(TacotronSTFT, self).__init__()
+        self.n_mel_channels = n_mel_channels
+        self.sampling_rate = sampling_rate
+        self.stft_fn = STFT(filter_length, hop_length, win_length)
+        mel_basis = librosa_mel_fn(sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax)
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+
+    def spectral_normalize(self, magnitudes, normalize_fun):
+        output = dynamic_range_compression(magnitudes, normalize_fun)
+        return output
+
+    def spectral_de_normalize(self, magnitudes):
+        output = dynamic_range_decompression(magnitudes)
+        return output
+
+    def mel_spectrogram(self, y, normalize_fun=torch.log):
+        """Computes mel-spectrograms from a batch of waves
+        PARAMS
+        ------
+        y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
+
+        RETURNS
+        -------
+        mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
+        """
+        assert torch.min(y.data) >= -1, torch.min(y.data)
+        assert torch.max(y.data) <= 1, torch.max(y.data)
+
+        magnitudes, phases = self.stft_fn.transform(y)
+        magnitudes = magnitudes.data
+        mel_output = torch.matmul(self.mel_basis, magnitudes)
+        mel_output = self.spectral_normalize(mel_output, normalize_fun)
+        energy = torch.norm(magnitudes, dim=1)
+
+        log_magnitudes = self.spectral_normalize(magnitudes, normalize_fun)
+
+        return mel_output, log_magnitudes, energy
+
+
+def pad_wav(waveform, segment_length):
+    waveform_length = waveform.shape[-1]
+    assert waveform_length > 100, "Waveform is too short, %s" % waveform_length
+    if segment_length is None or waveform_length == segment_length:
+        return waveform
+    elif waveform_length > segment_length:
+        return waveform[:, :segment_length]
+    elif waveform_length < segment_length:
+        temp_wav = np.zeros((1, segment_length))
+        temp_wav[:, :waveform_length] = waveform
+    return temp_wav
+
+
+def normalize_wav(waveform):
+    waveform = waveform - np.mean(waveform)
+    waveform = waveform / (np.max(np.abs(waveform)) + 1e-8)
+    return waveform * 0.5
+
+
+def _pad_spec(fbank, target_length=1024):
+    n_frames = fbank.shape[0]
+    p = target_length - n_frames
+    # cut and pad
+    if p > 0:
+        m = torch.nn.ZeroPad2d((0, 0, 0, p))
+        fbank = m(fbank)
+    elif p < 0:
+        fbank = fbank[0:target_length, :]
+
+    if fbank.size(-1) % 2 != 0:
+        fbank = fbank[..., :-1]
+
+    return fbank
+
+
+def get_mel_from_wav(audio, _stft):
+    audio = torch.clip(torch.FloatTensor(audio).unsqueeze(0), -1, 1)
+    audio = torch.autograd.Variable(audio, requires_grad=False)
+    melspec, log_magnitudes_stft, energy = _stft.mel_spectrogram(audio)
+    melspec = torch.squeeze(melspec, 0).numpy().astype(np.float32)
+    log_magnitudes_stft = torch.squeeze(log_magnitudes_stft, 0).numpy().astype(np.float32)
+    energy = torch.squeeze(energy, 0).numpy().astype(np.float32)
+    return melspec, log_magnitudes_stft, energy
+
+
+def read_wav_file_io(bytes):
+    # waveform, sr = librosa.load(filename, sr=None, mono=True) # 4 times slower
+    waveform, sr = torchaudio.load(bytes, format="mp4")  # Faster!!!
+    waveform = torchaudio.functional.resample(waveform, orig_freq=sr, new_freq=16000)
+    # waveform = waveform.numpy()[0, ...]
+    # waveform = normalize_wav(waveform)
+    # waveform = waveform[None, ...]
+
+    # waveform = waveform / (np.max(np.abs(waveform)) + 1e-8)
+    # waveform = 0.5 * waveform
+
+    return waveform
+
+
+def load_audio(bytes, sample_rate=16000):
+    waveform, sr = torchaudio.load(bytes, format="mp4")
+    waveform = torchaudio.functional.resample(waveform, orig_freq=sr, new_freq=sample_rate)
+    return waveform
+
+
+def read_wav_file(filename):
+    # waveform, sr = librosa.load(filename, sr=None, mono=True) # 4 times slower
+    waveform, sr = torchaudio.load(filename)  # Faster!!!
+    waveform = torchaudio.functional.resample(waveform, orig_freq=sr, new_freq=16000)
+    waveform = waveform.numpy()[0, ...]
+    waveform = normalize_wav(waveform)
+    waveform = waveform[None, ...]
+
+    waveform = waveform / np.max(np.abs(waveform))
+    waveform = 0.5 * waveform
+
+    return waveform
+
+
+def norm_wav_tensor(waveform: torch.FloatTensor):
+    waveform = waveform.numpy()[0, ...]
+    waveform = normalize_wav(waveform)
+    waveform = waveform[None, ...]
+    waveform = waveform / (np.max(np.abs(waveform)) + 1e-8)
+    waveform = 0.5 * waveform
+    return waveform
+
+
+def wav_to_fbank(filename, target_length=1024, fn_STFT=None):
+    if fn_STFT is None:
+        fn_STFT = TacotronSTFT(
+            1024,  # filter_length
+            160,  # hop_length
+            1024,  # win_length
+            64,  # n_mel
+            16000,  # sample_rate
+            0,  # fmin
+            8000,  # fmax
+        )
+
+    # mixup
+    waveform = read_wav_file(filename, target_length * 160)  # hop size is 160
+
+    waveform = waveform[0, ...]
+    waveform = torch.FloatTensor(waveform)
+
+    fbank, log_magnitudes_stft, energy = get_mel_from_wav(waveform, fn_STFT)
+
+    fbank = torch.FloatTensor(fbank.T)
+    log_magnitudes_stft = torch.FloatTensor(log_magnitudes_stft.T)
+
+    fbank, log_magnitudes_stft = _pad_spec(fbank, target_length), _pad_spec(log_magnitudes_stft, target_length)
+
+    return fbank, log_magnitudes_stft, waveform
+
+
+def wav_tensor_to_fbank(waveform, target_length=512, fn_STFT=None):
+    if fn_STFT is None:
+        fn_STFT = TacotronSTFT(
+            1024,  # filter_length
+            160,  # hop_length
+            1024,  # win_length
+            256,  # n_mel
+            16000,  # sample_rate
+            0,  # fmin
+            8000,  # fmax
+        )  # In practice used
+
+    fbank, log_magnitudes_stft, energy = get_mel_from_wav(waveform, fn_STFT)
+
+    fbank = torch.FloatTensor(fbank.T)
+    log_magnitudes_stft = torch.FloatTensor(log_magnitudes_stft.T)
+
+    fbank, log_magnitudes_stft = _pad_spec(fbank, target_length), _pad_spec(log_magnitudes_stft, target_length)
+
+    return fbank

requirements.txt

@@ -0,0 +1,6 @@
+#soundfile
+#safetensors
+#moviepy
+#decord
+#imageio
+#h5py