From fe06f8958aff3eca8098cc3539544c3582db0020 Mon Sep 17 00:00:00 2001 From: Bubbliiiing <47347516+bubbliiiing@users.noreply.github.com> Date: Wed, 22 Oct 2025 10:53:59 +0800 Subject: [PATCH] Update FantasyTalking training and Dataset Loading structure (#358) --- examples/fantasytalking/predict_s2v.py | 20 +- scripts/fantasytalking/README_TRAIN.md | 220 ++ scripts/fantasytalking/train.py | 1899 +++++++++++++++++ scripts/fantasytalking/train.sh | 40 + videox_fun/data/__init__.py | 9 + videox_fun/data/dataset_image.py | 2 +- videox_fun/data/dataset_image_video.py | 320 +-- videox_fun/data/dataset_video.py | 264 ++- videox_fun/data/utils.py | 347 +++ .../models/fantasytalking_audio_encoder.py | 9 + .../pipeline/pipeline_fantasy_talking.py | 2 +- 11 files changed, 2755 insertions(+), 377 deletions(-) create mode 100755 scripts/fantasytalking/README_TRAIN.md create mode 100644 scripts/fantasytalking/train.py create mode 100644 scripts/fantasytalking/train.sh create mode 100644 videox_fun/data/__init__.py create mode 100644 videox_fun/data/utils.py diff --git a/examples/fantasytalking/predict_s2v.py b/examples/fantasytalking/predict_s2v.py index 1150a31..ec8e5ed 100644 --- a/examples/fantasytalking/predict_s2v.py +++ b/examples/fantasytalking/predict_s2v.py @@ -141,14 +141,20 @@ if transformer_path is not None: else: state_dict = torch.load(transformer_path, map_location="cpu") - audio_processor_dict = state_dict["audio_processor"] if "audio_processor" in state_dict else state_dict - m, u = transformer.load_state_dict(audio_processor_dict, strict=False) - print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + if "audio_processor" in state_dict: + audio_processor_dict = state_dict["audio_processor"] if "audio_processor" in state_dict else state_dict + m, u = transformer.load_state_dict(audio_processor_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") - proj_model_dict = state_dict["proj_model"] if "proj_model" in state_dict else state_dict - proj_model_dict = {"proj_model." + k : v for k, v in proj_model_dict.items()} - m, u = transformer.load_state_dict(proj_model_dict, strict=False) - print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + proj_model_dict = state_dict["proj_model"] if "proj_model" in state_dict else state_dict + proj_model_dict = {"proj_model." + k : v for k, v in proj_model_dict.items()} + m, u = transformer.load_state_dict(proj_model_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + else: + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") # Get Vae Chosen_AutoencoderKL = { diff --git a/scripts/fantasytalking/README_TRAIN.md b/scripts/fantasytalking/README_TRAIN.md new file mode 100755 index 0000000..0edf4b7 --- /dev/null +++ b/scripts/fantasytalking/README_TRAIN.md @@ -0,0 +1,220 @@ +## Training Code + +The default training commands for the different versions are as follows: + +We can choose whether to use fsdp in FantasyTalking, which can save a lot of video memory. + +The metadata_control.json is a little different from normal json in FantasyTalking, you need to add a audio_path. + +```json +[ + { + "file_path": "train/00000001.mp4", + "audio_path": "wav/00000001.wav", + "text": "A group of young men in suits and sunglasses are walking down a city street.", + "type": "video" + }, + ..... +] +``` + +Some parameters in the sh file can be confusing, and they are explained in this document: + +- `enable_bucket` is used to enable bucket training. When enabled, the model does not crop the videos at the center, but instead, it trains the videos after grouping them into buckets based on resolution. +- `random_frame_crop` is used for random cropping on video frames to simulate videos with different frame counts. +- `random_hw_adapt` is used to enable automatic height and width scaling for videos. When `random_hw_adapt` is enabled, for training videos, the height and width will be set to `video_sample_size` as the maximum and `512` as the minimum. + - For example, when `random_hw_adapt` is enabled, with `video_sample_n_frames=49`, `video_sample_size=512`, the resolution of video inputs for training is `512x512x49`. +- `training_with_video_token_length` specifies training the model according to token length. For training videos, the height and width will be set to `video_sample_size` as the maximum and `256` as the minimum. + - For example, when `training_with_video_token_length` is enabled, with `video_sample_n_frames=49`, `token_sample_size=512`, `video_sample_size=256`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49` to `512x512x21`. + - The token length for a video with dimensions 512x512 and 49 frames is 13,312. We need to set the `token_sample_size = 512`. + - At 512x512 resolution, the number of video frames is 49 (~= 512 * 512 * 49 / 512 / 512). + - At 768x768 resolution, the number of video frames is 21 (~= 512 * 512 * 49 / 768 / 768). + - At 1024x1024 resolution, the number of video frames is 9 (~= 512 * 512 * 49 / 1024 / 1024). + - These resolutions combined with their corresponding lengths allow the model to generate videos of different sizes. +- `resume_from_checkpoint` is used to set the training should be resumed from a previous checkpoint. Use a path or `"latest"` to automatically select the last available checkpoint. + +FantasyTalking without deepspeed: + +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-I2V-14B-720P" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" scripts/fantasytalking/train.py \ + --config_path="config/wan2.1/wan_civitai.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --video_sample_size=512 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=2e-05 \ + --lr_scheduler="constant_with_warmup" \ + --lr_warmup_steps=100 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --low_vram \ + --transformer_path="models/FantasyTalking/fantasytalking_model.ckpt" \ + --trainable_modules "processor." "proj_model." +``` + +FantasyTalking with deepspeed zero-2: + +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-I2V-14B-720P" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/fantasytalking/train.py \ + --config_path="config/wan2.1/wan_civitai.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --video_sample_size=512 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=2e-05 \ + --lr_scheduler="constant_with_warmup" \ + --lr_warmup_steps=100 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --low_vram \ + --transformer_path="models/FantasyTalking/fantasytalking_model.ckpt" \ + --trainable_modules "processor." "proj_model." +``` + +FantasyTalking with deepspeed zero-3: + +```sh +python scripts/zero_to_bf16.py output_dir/checkpoint-{our-num-steps} output_dir/checkpoint-{your-num-steps}-outputs --max_shard_size 80GB --safe_serialization +``` + +Training shell command is as follows: +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-I2V-14B-720P" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/fantasytalking/train.py \ + --config_path="config/wan2.1/wan_civitai.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --video_sample_size=512 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=2e-05 \ + --lr_scheduler="constant_with_warmup" \ + --lr_warmup_steps=100 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --low_vram \ + --transformer_path="models/FantasyTalking/fantasytalking_model.ckpt" \ + --trainable_modules "processor." "proj_model." +``` + +FantasyTalking with FSDP: + +Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows: +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-I2V-14B-720P" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=AudioAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/fantasytalking/train.py \ + --config_path="config/wan2.1/wan_civitai.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --video_sample_size=512 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=2e-05 \ + --lr_scheduler="constant_with_warmup" \ + --lr_warmup_steps=100 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --low_vram \ + --transformer_path="models/FantasyTalking/fantasytalking_model.ckpt" \ + --trainable_modules "processor." "proj_model." +``` \ No newline at end of file diff --git a/scripts/fantasytalking/train.py b/scripts/fantasytalking/train.py new file mode 100644 index 0000000..668a180 --- /dev/null +++ b/scripts/fantasytalking/train.py @@ -0,0 +1,1899 @@ +"""Modified from https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py +""" +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 The HuggingFace Inc. 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 + +import argparse +import gc +import logging +import math +import os +import pickle +import random +import shutil +import sys + +import accelerate +import diffusers +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +import torchvision.transforms.functional as TF +import transformers +from accelerate import Accelerator, FullyShardedDataParallelPlugin +from accelerate.logging import get_logger +from accelerate.state import AcceleratorState +from accelerate.utils import ProjectConfiguration, set_seed +from diffusers import DDIMScheduler, FlowMatchEulerDiscreteScheduler +from diffusers.optimization import get_scheduler +from diffusers.training_utils import (EMAModel, + compute_density_for_timestep_sampling, + compute_loss_weighting_for_sd3) +from diffusers.utils import check_min_version, deprecate, is_wandb_available +from diffusers.utils.torch_utils import is_compiled_module +from einops import rearrange +from omegaconf import OmegaConf +from packaging import version +from PIL import Image +from torch.distributed.fsdp.fully_sharded_data_parallel import ( + FullOptimStateDictConfig, FullStateDictConfig, ShardedOptimStateDictConfig, + ShardedStateDictConfig) +from torch.utils.data import RandomSampler +from torch.utils.tensorboard import SummaryWriter +from torchvision import transforms +from tqdm.auto import tqdm +from transformers import AutoTokenizer +from transformers.utils import ContextManagers + +import datasets + +current_file_path = os.path.abspath(__file__) +project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))] +for project_root in project_roots: + sys.path.insert(0, project_root) if project_root not in sys.path else None + +from videox_fun.data.bucket_sampler import (ASPECT_RATIO_512, + ASPECT_RATIO_RANDOM_CROP_512, + ASPECT_RATIO_RANDOM_CROP_PROB, + AspectRatioBatchImageVideoSampler, + RandomSampler, get_closest_ratio) +from videox_fun.data.dataset_image_video import ImageVideoSampler, get_random_mask +from videox_fun.data.dataset_video import VideoSpeechDataset +from videox_fun.models import (AutoencoderKLWan, CLIPModel, WanT5EncoderModel, FantasyTalkingAudioEncoder, + FantasyTalkingTransformer3DModel) +from videox_fun.pipeline import FantasyTalkingPipeline, WanFunPipeline +from videox_fun.utils.discrete_sampler import DiscreteSampling +from videox_fun.utils.utils import get_image_to_video_latent, save_videos_grid + +if is_wandb_available(): + import wandb + + +def filter_kwargs(cls, kwargs): + import inspect + sig = inspect.signature(cls.__init__) + valid_params = set(sig.parameters.keys()) - {'self', 'cls'} + filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_params} + return filtered_kwargs + +def resize_mask(mask, latent, process_first_frame_only=True): + latent_size = latent.size() + batch_size, channels, num_frames, height, width = mask.shape + + if process_first_frame_only: + target_size = list(latent_size[2:]) + target_size[0] = 1 + first_frame_resized = F.interpolate( + mask[:, :, 0:1, :, :], + size=target_size, + mode='trilinear', + align_corners=False + ) + + target_size = list(latent_size[2:]) + target_size[0] = target_size[0] - 1 + if target_size[0] != 0: + remaining_frames_resized = F.interpolate( + mask[:, :, 1:, :, :], + size=target_size, + mode='trilinear', + align_corners=False + ) + resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2) + else: + resized_mask = first_frame_resized + else: + target_size = list(latent_size[2:]) + resized_mask = F.interpolate( + mask, + size=target_size, + mode='trilinear', + align_corners=False + ) + return resized_mask + +def linear_decay(initial_value, final_value, total_steps, current_step): + if current_step >= total_steps: + return final_value + current_step = max(0, current_step) + step_size = (final_value - initial_value) / total_steps + current_value = initial_value + step_size * current_step + return current_value + +def generate_timestep_with_lognorm(low, high, shape, device="cpu", generator=None): + u = torch.normal(mean=0.0, std=1.0, size=shape, device=device, generator=generator) + t = 1 / (1 + torch.exp(-u)) * (high - low) + low + return torch.clip(t.to(torch.int32), low, high - 1) + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.18.0.dev0") + +logger = get_logger(__name__, log_level="INFO") + +def log_validation(vae, text_encoder, tokenizer, clip_image_encoder, transformer3d, args, config, accelerator, weight_dtype, global_step): + try: + logger.info("Running validation... ") + + transformer3d_val = FantasyTalkingTransformer3DModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['transformer_additional_kwargs'].get('transformer_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + ).to(weight_dtype) + transformer3d_val.load_state_dict(accelerator.unwrap_model(transformer3d).state_dict()) + scheduler = FlowMatchEulerDiscreteScheduler( + **filter_kwargs(FlowMatchEulerDiscreteScheduler, OmegaConf.to_container(config['scheduler_kwargs'])) + ) + + pipeline = FantasyTalkingPipeline( + vae=accelerator.unwrap_model(vae).to(weight_dtype), + text_encoder=accelerator.unwrap_model(text_encoder), + tokenizer=tokenizer, + transformer=transformer3d_val, + scheduler=scheduler, + clip_image_encoder=clip_image_encoder, + ) + pipeline = pipeline.to(accelerator.device) + + if args.seed is None: + generator = None + else: + generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) + + images = [] + for i in range(len(args.validation_prompts)): + with torch.no_grad(): + with torch.autocast("cuda", dtype=weight_dtype): + video_length = int(args.video_sample_n_frames // vae.config.temporal_compression_ratio * vae.config.temporal_compression_ratio) + 1 if args.video_sample_n_frames != 1 else 1 + input_video, input_video_mask, clip_image = get_image_to_video_latent(args.validation_image_paths[i], None, video_length=video_length, sample_size=[args.video_sample_size, args.video_sample_size]) + audio_path = args.validation_audio_paths[i] + + sample = pipeline( + args.validation_prompts[i], + num_frames = args.video_sample_n_frames, + negative_prompt = "bad detailed", + height = args.video_sample_size, + width = args.video_sample_size, + generator = generator, + + video = input_video, + mask_video = input_video_mask, + clip_image = clip_image, + audio_path = audio_path, + shift = 5, + fps = 16 + ).videos + os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True) + save_videos_grid(sample, os.path.join(args.output_dir, f"sample/sample-{global_step}-{i}.gif")) + + del pipeline + del transformer3d_val + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + + return images + except Exception as e: + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + print(f"Eval error with info {e}") + return None + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1." + ) + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + required=True, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--revision", + type=str, + default=None, + required=False, + help="Revision of pretrained model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--variant", + type=str, + default=None, + help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. " + ), + ) + parser.add_argument( + "--train_data_meta", + type=str, + default=None, + help=( + "A csv containing the training data. " + ), + ) + parser.add_argument( + "--max_train_samples", + type=int, + default=None, + help=( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ), + ) + parser.add_argument( + "--validation_prompts", + type=str, + default=None, + nargs="+", + help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."), + ) + parser.add_argument( + "--validation_image_paths", + type=str, + default=None, + nargs="+", + help=("A set of images evaluated every `--validation_epochs` and logged to `--report_to`."), + ) + parser.add_argument( + "--validation_audio_paths", + type=str, + default=None, + nargs="+", + help=("A set of audios evaluated every `--validation_epochs` and logged to `--report_to`."), + ) + parser.add_argument( + "--output_dir", + type=str, + default="sd-model-finetuned", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--random_flip", + action="store_true", + help="whether to randomly flip images horizontally", + ) + parser.add_argument( + "--use_came", + action="store_true", + help="whether to use came", + ) + parser.add_argument( + "--multi_stream", + action="store_true", + help="whether to use cuda multi-stream", + ) + parser.add_argument( + "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument( + "--vae_mini_batch", type=int, default=32, help="mini batch size for vae." + ) + parser.add_argument("--num_train_epochs", type=int, default=100) + parser.add_argument( + "--max_train_steps", + type=int, + default=None, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=1e-4, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.") + parser.add_argument( + "--non_ema_revision", + type=str, + default=None, + required=False, + help=( + "Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or" + " remote repository specified with --pretrained_model_name_or_path." + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--prediction_type", + type=str, + default=None, + help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.", + ) + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--report_model_info", action="store_true", help="Whether or not to report more info about model (such as norm, grad)." + ) + parser.add_argument( + "--mixed_precision", + type=str, + default=None, + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" + " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" + " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." + ), + ) + parser.add_argument( + "--report_to", + type=str, + default="tensorboard", + help=( + 'The integration to report the results and logs to. Supported platforms are `"tensorboard"`' + ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.' + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming" + " training using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=("Max number of checkpoints to store."), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.") + parser.add_argument( + "--validation_epochs", + type=int, + default=5, + help="Run validation every X epochs.", + ) + parser.add_argument( + "--validation_steps", + type=int, + default=2000, + help="Run validation every X steps.", + ) + parser.add_argument( + "--tracker_project_name", + type=str, + default="text2image-fine-tune", + help=( + "The `project_name` argument passed to Accelerator.init_trackers for" + " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator" + ), + ) + + parser.add_argument( + "--snr_loss", action="store_true", help="Whether or not to use snr_loss." + ) + parser.add_argument( + "--uniform_sampling", action="store_true", help="Whether or not to use uniform_sampling." + ) + parser.add_argument( + "--enable_text_encoder_in_dataloader", action="store_true", help="Whether or not to use text encoder in dataloader." + ) + parser.add_argument( + "--enable_bucket", action="store_true", help="Whether enable bucket sample in datasets." + ) + parser.add_argument( + "--random_ratio_crop", action="store_true", help="Whether enable random ratio crop sample in datasets." + ) + parser.add_argument( + "--random_frame_crop", action="store_true", help="Whether enable random frame crop sample in datasets." + ) + parser.add_argument( + "--random_hw_adapt", action="store_true", help="Whether enable random adapt height and width in datasets." + ) + parser.add_argument( + "--training_with_video_token_length", action="store_true", help="The training stage of the model in training.", + ) + parser.add_argument( + "--auto_tile_batch_size", action="store_true", help="Whether to auto tile batch size.", + ) + parser.add_argument( + "--motion_sub_loss", action="store_true", help="Whether enable motion sub loss." + ) + parser.add_argument( + "--motion_sub_loss_ratio", type=float, default=0.25, help="The ratio of motion sub loss." + ) + parser.add_argument( + "--train_sampling_steps", + type=int, + default=1000, + help="Run train_sampling_steps.", + ) + parser.add_argument( + "--keep_all_node_same_token_length", + action="store_true", + help="Reference of the length token.", + ) + parser.add_argument( + "--token_sample_size", + type=int, + default=512, + help="Sample size of the token.", + ) + parser.add_argument( + "--video_sample_size", + type=int, + default=512, + help="Sample size of the video.", + ) + parser.add_argument( + "--fix_sample_size", + nargs=2, type=int, default=None, + help="Fix Sample size [height, width] when using bucket and collate_fn." + ) + parser.add_argument( + "--video_sample_stride", + type=int, + default=4, + help="Sample stride of the video.", + ) + parser.add_argument( + "--video_sample_n_frames", + type=int, + default=17, + help="Num frame of video.", + ) + parser.add_argument( + "--config_path", + type=str, + default=None, + help=( + "The config of the model in training." + ), + ) + parser.add_argument( + "--transformer_path", + type=str, + default=None, + help=("If you want to load the weight from other transformers, input its path."), + ) + parser.add_argument( + "--vae_path", + type=str, + default=None, + help=("If you want to load the weight from other vaes, input its path."), + ) + + parser.add_argument( + '--trainable_modules', + nargs='+', + help='Enter a list of trainable modules' + ) + parser.add_argument( + '--trainable_modules_low_learning_rate', + nargs='+', + default=[], + help='Enter a list of trainable modules with lower learning rate' + ) + parser.add_argument( + '--tokenizer_max_length', + type=int, + default=512, + help='Max length of tokenizer' + ) + parser.add_argument( + "--use_deepspeed", action="store_true", help="Whether or not to use deepspeed." + ) + parser.add_argument( + "--use_fsdp", action="store_true", help="Whether or not to use fsdp." + ) + parser.add_argument( + "--low_vram", action="store_true", help="Whether enable low_vram mode." + ) + parser.add_argument( + "--abnormal_norm_clip_start", + type=int, + default=1000, + help=( + 'When do we start doing additional processing on abnormal gradients. ' + ), + ) + parser.add_argument( + "--initial_grad_norm_ratio", + type=int, + default=5, + help=( + 'The initial gradient is relative to the multiple of the max_grad_norm. ' + ), + ) + parser.add_argument( + "--weighting_scheme", + type=str, + default="none", + choices=["sigma_sqrt", "logit_normal", "mode", "cosmap", "none"], + help=('We default to the "none" weighting scheme for uniform sampling and uniform loss'), + ) + parser.add_argument( + "--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--mode_scale", + type=float, + default=1.29, + help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.", + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # default to using the same revision for the non-ema model if not specified + if args.non_ema_revision is None: + args.non_ema_revision = args.revision + + return args + + +def main(): + args = parse_args() + + if args.report_to == "wandb" and args.hub_token is not None: + raise ValueError( + "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token." + " Please use `huggingface-cli login` to authenticate with the Hub." + ) + + if args.non_ema_revision is not None: + deprecate( + "non_ema_revision!=None", + "0.15.0", + message=( + "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to" + " use `--variant=non_ema` instead." + ), + ) + logging_dir = os.path.join(args.output_dir, args.logging_dir) + + config = OmegaConf.load(args.config_path) + accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) + + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with=args.report_to, + project_config=accelerator_project_config, + ) + + deepspeed_plugin = accelerator.state.deepspeed_plugin if hasattr(accelerator.state, "deepspeed_plugin") else None + fsdp_plugin = accelerator.state.fsdp_plugin if hasattr(accelerator.state, "fsdp_plugin") else None + if deepspeed_plugin is not None: + zero_stage = int(deepspeed_plugin.zero_stage) + fsdp_stage = 0 + print(f"Using DeepSpeed Zero stage: {zero_stage}") + + args.use_deepspeed = True + if zero_stage == 3: + print(f"Auto set save_state to True because zero_stage == 3") + args.save_state = True + elif fsdp_plugin is not None: + from torch.distributed.fsdp import ShardingStrategy + zero_stage = 0 + if fsdp_plugin.sharding_strategy is ShardingStrategy.FULL_SHARD: + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is None: # The fsdp_plugin.sharding_strategy is None in FSDP 2. + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is ShardingStrategy.SHARD_GRAD_OP: + fsdp_stage = 2 + else: + fsdp_stage = 0 + print(f"Using FSDP stage: {fsdp_stage}") + + args.use_fsdp = True + if fsdp_stage == 3: + print(f"Auto set save_state to True because fsdp_stage == 3") + args.save_state = True + else: + zero_stage = 0 + fsdp_stage = 0 + print("DeepSpeed is not enabled.") + + if accelerator.is_main_process: + writer = SummaryWriter(log_dir=logging_dir) + + # Make one log on every process with the configuration for debugging. + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + level=logging.INFO, + ) + logger.info(accelerator.state, main_process_only=False) + if accelerator.is_local_main_process: + datasets.utils.logging.set_verbosity_warning() + transformers.utils.logging.set_verbosity_warning() + diffusers.utils.logging.set_verbosity_info() + else: + datasets.utils.logging.set_verbosity_error() + transformers.utils.logging.set_verbosity_error() + diffusers.utils.logging.set_verbosity_error() + + # If passed along, set the training seed now. + if args.seed is not None: + set_seed(args.seed) + rng = np.random.default_rng(np.random.PCG64(args.seed + accelerator.process_index)) + torch_rng = torch.Generator(accelerator.device).manual_seed(args.seed + accelerator.process_index) + else: + rng = None + torch_rng = None + index_rng = np.random.default_rng(np.random.PCG64(43)) + print(f"Init rng with seed {args.seed + accelerator.process_index}. Process_index is {accelerator.process_index}") + + # Handle the repository creation + if accelerator.is_main_process: + if args.output_dir is not None: + os.makedirs(args.output_dir, exist_ok=True) + + # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora transformer3d) to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + args.mixed_precision = accelerator.mixed_precision + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + args.mixed_precision = accelerator.mixed_precision + + # Load scheduler, tokenizer and models. + noise_scheduler = FlowMatchEulerDiscreteScheduler( + **filter_kwargs(FlowMatchEulerDiscreteScheduler, OmegaConf.to_container(config['scheduler_kwargs'])) + ) + + # Get Tokenizer + tokenizer = AutoTokenizer.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['text_encoder_kwargs'].get('tokenizer_subpath', 'tokenizer')), + ) + + def deepspeed_zero_init_disabled_context_manager(): + """ + returns either a context list that includes one that will disable zero.Init or an empty context list + """ + deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None + if deepspeed_plugin is None: + return [] + + return [deepspeed_plugin.zero3_init_context_manager(enable=False)] + + # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3. + # For this to work properly all models must be run through `accelerate.prepare`. But accelerate + # will try to assign the same optimizer with the same weights to all models during + # `deepspeed.initialize`, which of course doesn't work. + # + # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2 + # frozen models from being partitioned during `zero.Init` which gets called during + # `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding + # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded. + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + # Get Text encoder + text_encoder = WanT5EncoderModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['text_encoder_kwargs'].get('text_encoder_subpath', 'text_encoder')), + additional_kwargs=OmegaConf.to_container(config['text_encoder_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, + ) + text_encoder = text_encoder.eval() + # Get Vae + vae = AutoencoderKLWan.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['vae_kwargs'].get('vae_subpath', 'vae')), + additional_kwargs=OmegaConf.to_container(config['vae_kwargs']), + ) + vae = vae.eval() + # Get Clip Image Encoder + clip_image_encoder = CLIPModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['image_encoder_kwargs'].get('image_encoder_subpath', 'image_encoder')), + ) + clip_image_encoder = clip_image_encoder.eval() + + audio_encoder = FantasyTalkingAudioEncoder( + os.path.join(args.pretrained_model_name_or_path, "audio_encoder") + ) + + # Get Transformer + transformer3d = FantasyTalkingTransformer3DModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['transformer_additional_kwargs'].get('transformer_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + ).to(weight_dtype) + + # Freeze vae and text_encoder and set transformer3d to trainable + vae.requires_grad_(False) + text_encoder.requires_grad_(False) + transformer3d.requires_grad_(False) + clip_image_encoder.requires_grad_(False) + audio_encoder.requires_grad_(False) + + if args.transformer_path is not None: + print(f"From checkpoint: {args.transformer_path}") + if args.transformer_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.transformer_path) + else: + state_dict = torch.load(args.transformer_path, map_location="cpu") + + if "audio_processor" in state_dict: + audio_processor_dict = state_dict["audio_processor"] if "audio_processor" in state_dict else state_dict + m, u = transformer3d.load_state_dict(audio_processor_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + + proj_model_dict = state_dict["proj_model"] if "proj_model" in state_dict else state_dict + proj_model_dict = {"proj_model." + k : v for k, v in proj_model_dict.items()} + m, u = transformer3d.load_state_dict(proj_model_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + else: + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + + if args.vae_path is not None: + print(f"From checkpoint: {args.vae_path}") + if args.vae_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.vae_path) + else: + state_dict = torch.load(args.vae_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = vae.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + assert len(u) == 0 + + # A good trainable modules is showed below now. + # For 3D Patch: trainable_modules = ['ff.net', 'pos_embed', 'attn2', 'proj_out', 'timepositionalencoding', 'h_position', 'w_position'] + # For 2D Patch: trainable_modules = ['ff.net', 'attn2', 'timepositionalencoding', 'h_position', 'w_position'] + transformer3d.train() + if accelerator.is_main_process: + accelerator.print( + f"Trainable modules '{args.trainable_modules}'." + ) + for name, param in transformer3d.named_parameters(): + for trainable_module_name in args.trainable_modules + args.trainable_modules_low_learning_rate: + if trainable_module_name in name: + param.requires_grad = True + break + + # Create EMA for the transformer3d. + if args.use_ema: + if zero_stage == 3: + raise NotImplementedError("FSDP does not support EMA.") + + ema_transformer3d = FantasyTalkingTransformer3DModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['transformer_additional_kwargs'].get('transformer_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + ).to(weight_dtype) + + ema_transformer3d = EMAModel(ema_transformer3d.parameters(), model_cls=FantasyTalkingTransformer3DModel, model_config=ema_transformer3d.config) + + # `accelerate` 0.16.0 will have better support for customized saving + if version.parse(accelerate.__version__) >= version.parse("0.16.0"): + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + if fsdp_stage != 0: + def save_model_hook(models, weights, output_dir): + accelerate_state_dict = accelerator.get_state_dict(models[-1], unwrap=True) + if accelerator.is_main_process: + from safetensors.torch import save_file + + safetensor_save_path = os.path.join(output_dir, f"diffusion_pytorch_model.safetensors") + accelerate_state_dict = {k: v.to(dtype=weight_dtype) for k, v in accelerate_state_dict.items()} + save_file(accelerate_state_dict, safetensor_save_path, metadata={"format": "pt"}) + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + elif zero_stage == 3: + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + else: + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + if args.use_ema: + ema_transformer3d.save_pretrained(os.path.join(output_dir, "transformer_ema")) + + models[0].save_pretrained(os.path.join(output_dir, "transformer")) + if not args.use_deepspeed: + weights.pop() + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + if args.use_ema: + ema_path = os.path.join(input_dir, "transformer_ema") + _, ema_kwargs = FantasyTalkingTransformer3DModel.load_config(ema_path, return_unused_kwargs=True) + load_model = FantasyTalkingTransformer3DModel.from_pretrained( + input_dir, subfolder="transformer_ema", + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']) + ) + load_model = EMAModel(load_model.parameters(), model_cls=FantasyTalkingTransformer3DModel, model_config=load_model.config) + load_model.load_state_dict(ema_kwargs) + + ema_transformer3d.load_state_dict(load_model.state_dict()) + ema_transformer3d.to(accelerator.device) + del load_model + + for i in range(len(models)): + # pop models so that they are not loaded again + model = models.pop() + + # load diffusers style into model + load_model = FantasyTalkingTransformer3DModel.from_pretrained( + input_dir, subfolder="transformer" + ) + model.register_to_config(**load_model.config) + + model.load_state_dict(load_model.state_dict()) + del load_model + + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + accelerator.register_save_state_pre_hook(save_model_hook) + accelerator.register_load_state_pre_hook(load_model_hook) + + if args.gradient_checkpointing: + transformer3d.enable_gradient_checkpointing() + + # Enable TF32 for faster training on Ampere GPUs, + # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Initialize the optimizer + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + elif args.use_came: + try: + from came_pytorch import CAME + except: + raise ImportError( + "Please install came_pytorch to use CAME. You can do so by running `pip install came_pytorch`" + ) + + optimizer_cls = CAME + else: + optimizer_cls = torch.optim.AdamW + + trainable_params = list(filter(lambda p: p.requires_grad, transformer3d.parameters())) + trainable_params_optim = [ + {'params': [], 'lr': args.learning_rate}, + {'params': [], 'lr': args.learning_rate / 2}, + ] + in_already = [] + for name, param in transformer3d.named_parameters(): + high_lr_flag = False + if name in in_already: + continue + for trainable_module_name in args.trainable_modules: + if trainable_module_name in name: + in_already.append(name) + high_lr_flag = True + trainable_params_optim[0]['params'].append(param) + if accelerator.is_main_process: + print(f"Set {name} to lr : {args.learning_rate}") + break + if high_lr_flag: + continue + for trainable_module_name in args.trainable_modules_low_learning_rate: + if trainable_module_name in name: + in_already.append(name) + trainable_params_optim[1]['params'].append(param) + if accelerator.is_main_process: + print(f"Set {name} to lr : {args.learning_rate / 2}") + break + + if args.use_came: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + # weight_decay=args.adam_weight_decay, + betas=(0.9, 0.999, 0.9999), + eps=(1e-30, 1e-16) + ) + else: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + # Get the training dataset + sample_n_frames_bucket_interval = vae.config.temporal_compression_ratio + + if args.fix_sample_size is not None and args.enable_bucket: + args.video_sample_size = max(max(args.fix_sample_size), args.video_sample_size) + args.training_with_video_token_length = False + args.random_hw_adapt = False + + # Get the dataset + train_dataset = VideoSpeechDataset( + args.train_data_meta, args.train_data_dir, + video_sample_size=args.video_sample_size, video_sample_stride=args.video_sample_stride, video_sample_n_frames=args.video_sample_n_frames, + enable_bucket=args.enable_bucket, enable_inpaint=True, + ) + + def worker_init_fn(_seed): + _seed = _seed * 256 + def _worker_init_fn(worker_id): + print(f"worker_init_fn with {_seed + worker_id}") + np.random.seed(_seed + worker_id) + random.seed(_seed + worker_id) + return _worker_init_fn + + if args.enable_bucket: + aspect_ratio_sample_size = {key : [x / 512 * args.video_sample_size for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = AspectRatioBatchImageVideoSampler( + sampler=RandomSampler(train_dataset, generator=batch_sampler_generator), dataset=train_dataset.dataset, + batch_size=args.train_batch_size, train_folder = args.train_data_dir, drop_last=True, + aspect_ratios=aspect_ratio_sample_size, + ) + + def collate_fn(examples): + def get_length_to_frame_num(token_length): + if args.video_sample_size > 256: + sample_sizes = list(range(256, args.video_sample_size + 1, 128)) + + if sample_sizes[-1] != args.video_sample_size: + sample_sizes.append(args.video_sample_size) + else: + sample_sizes = [args.video_sample_size] + + length_to_frame_num = { + sample_size: min(token_length / sample_size / sample_size, args.video_sample_n_frames) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1 for sample_size in sample_sizes + } + + return length_to_frame_num + + def get_random_downsample_ratio(sample_size, image_ratio=[], + all_choices=False, rng=None): + def _create_special_list(length): + if length == 1: + return [1.0] + if length >= 2: + first_element = 0.90 + remaining_sum = 1.0 - first_element + other_elements_value = remaining_sum / (length - 1) + special_list = [first_element] + [other_elements_value] * (length - 1) + return special_list + + if sample_size >= 1536: + number_list = [1, 1.25, 1.5, 2, 2.5, 3] + image_ratio + elif sample_size >= 1024: + number_list = [1, 1.25, 1.5, 2] + image_ratio + elif sample_size >= 768: + number_list = [1, 1.25, 1.5] + image_ratio + elif sample_size >= 512: + number_list = [1] + image_ratio + else: + number_list = [1] + + if all_choices: + return number_list + + number_list_prob = np.array(_create_special_list(len(number_list))) + if rng is None: + return np.random.choice(number_list, p = number_list_prob) + else: + return rng.choice(number_list, p = number_list_prob) + + # Get token length + target_token_length = args.video_sample_n_frames * args.token_sample_size * args.token_sample_size + length_to_frame_num = get_length_to_frame_num(target_token_length) + + # Create new output + new_examples = {} + new_examples["target_token_length"] = target_token_length + new_examples["pixel_values"] = [] + new_examples["text"] = [] + new_examples["audio"] = [] + new_examples["sample_rate"] = [] + + # Used in Inpaint mode + new_examples["mask_pixel_values"] = [] + new_examples["mask"] = [] + new_examples["clip_pixel_values"] = [] + + # Get downsample ratio in image and videos + pixel_value = examples[0]["pixel_values"] + f, h, w, c = np.shape(pixel_value) + + if args.random_hw_adapt: + if args.training_with_video_token_length: + local_min_size = np.min(np.array([np.mean(np.array([np.shape(example["pixel_values"])[1], np.shape(example["pixel_values"])[2]])) for example in examples])) + # The video will be resized to a lower resolution than its own. + choice_list = [length for length in list(length_to_frame_num.keys()) if length < local_min_size * 1.25] + if len(choice_list) == 0: + choice_list = list(length_to_frame_num.keys()) + local_video_sample_size = np.random.choice(choice_list) + batch_video_length = length_to_frame_num[local_video_sample_size] + random_downsample_ratio = args.video_sample_size / local_video_sample_size + else: + random_downsample_ratio = get_random_downsample_ratio(args.video_sample_size) + batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval + else: + random_downsample_ratio = 1 + batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval + + aspect_ratio_sample_size = {key : [x / 512 * args.video_sample_size / random_downsample_ratio for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + aspect_ratio_random_crop_sample_size = {key : [x / 512 * args.video_sample_size / random_downsample_ratio for x in ASPECT_RATIO_RANDOM_CROP_512[key]] for key in ASPECT_RATIO_RANDOM_CROP_512.keys()} + + if args.fix_sample_size is not None: + fix_sample_size = [int(x / 16) * 16 for x in args.fix_sample_size] + elif args.random_ratio_crop: + if rng is None: + random_sample_size = aspect_ratio_random_crop_sample_size[ + np.random.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + else: + random_sample_size = aspect_ratio_random_crop_sample_size[ + rng.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + random_sample_size = [int(x / 16) * 16 for x in random_sample_size] + else: + closest_size, closest_ratio = get_closest_ratio(h, w, ratios=aspect_ratio_sample_size) + closest_size = [int(x / 16) * 16 for x in closest_size] + + min_example_length = min( + [example["pixel_values"].shape[0] for example in examples] + ) + batch_video_length = int(min(batch_video_length, min_example_length)) + + # Magvae needs the number of frames to be 4n + 1. + batch_video_length = (batch_video_length - 1) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1 + + if batch_video_length <= 0: + batch_video_length = 1 + + for example in examples: + if args.fix_sample_size is not None: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + fix_sample_size = list(map(lambda x: int(x), fix_sample_size)) + transform = transforms.Compose([ + transforms.Resize(fix_sample_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(fix_sample_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + elif args.random_ratio_crop: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + b, c, h, w = pixel_values.size() + th, tw = random_sample_size + if th / tw > h / w: + nh = int(th) + nw = int(w / h * nh) + else: + nw = int(tw) + nh = int(h / w * nw) + + transform = transforms.Compose([ + transforms.Resize([nh, nw]), + transforms.CenterCrop([int(x) for x in random_sample_size]), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + else: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + closest_size = list(map(lambda x: int(x), closest_size)) + if closest_size[0] / h > closest_size[1] / w: + resize_size = closest_size[0], int(w * closest_size[0] / h) + else: + resize_size = int(h * closest_size[1] / w), closest_size[1] + + transform = transforms.Compose([ + transforms.Resize(resize_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(closest_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + + new_examples["pixel_values"].append(transform(pixel_values)[:batch_video_length]) + new_examples["text"].append(example["text"]) + + audio_length = np.shape(example["audio"])[0] + batch_audio_length = int(audio_length / pixel_values.size()[0] * batch_video_length) + new_examples["audio"].append(example["audio"][:batch_audio_length]) + new_examples["sample_rate"].append(example["sample_rate"]) + + mask = get_random_mask(new_examples["pixel_values"][-1].size(), image_start_only=True) + mask_pixel_values = new_examples["pixel_values"][-1] * (1 - mask) + # Wan 2.1 use 0 for masked pixels + # + torch.ones_like(new_examples["pixel_values"][-1]) * -1 * mask + new_examples["mask_pixel_values"].append(mask_pixel_values) + new_examples["mask"].append(mask) + + clip_pixel_values = new_examples["pixel_values"][-1][0].permute(1, 2, 0).contiguous() + clip_pixel_values = (clip_pixel_values * 0.5 + 0.5) * 255 + new_examples["clip_pixel_values"].append(clip_pixel_values) + + # Limit the number of frames to the same + new_examples["pixel_values"] = torch.stack([example for example in new_examples["pixel_values"]]) + new_examples["mask_pixel_values"] = torch.stack([example for example in new_examples["mask_pixel_values"]]) + new_examples["mask"] = torch.stack([example for example in new_examples["mask"]]) + new_examples["clip_pixel_values"] = torch.stack([example for example in new_examples["clip_pixel_values"]]) + new_examples["audio"] = torch.stack([example for example in new_examples["audio"]]) + + # Encode prompts when enable_text_encoder_in_dataloader=True + if args.enable_text_encoder_in_dataloader: + prompt_ids = tokenizer( + new_examples['text'], + max_length=args.tokenizer_max_length, + padding="max_length", + add_special_tokens=True, + truncation=True, + return_tensors="pt" + ) + encoder_hidden_states = text_encoder( + prompt_ids.input_ids + )[0] + new_examples['encoder_attention_mask'] = prompt_ids.attention_mask + new_examples['encoder_hidden_states'] = encoder_hidden_states + + return new_examples + + # DataLoaders creation: + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + collate_fn=collate_fn, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + worker_init_fn=worker_init_fn(args.seed + accelerator.process_index) + ) + else: + # DataLoaders creation: + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = ImageVideoSampler(RandomSampler(train_dataset, generator=batch_sampler_generator), train_dataset, args.train_batch_size) + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + worker_init_fn=worker_init_fn(args.seed + accelerator.process_index) + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes, + num_training_steps=args.max_train_steps * accelerator.num_processes, + ) + + # Prepare everything with our `accelerator`. + transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + transformer3d, optimizer, train_dataloader, lr_scheduler + ) + + if fsdp_stage != 0: + from functools import partial + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype) + text_encoder = shard_fn(text_encoder) + + if args.use_ema: + ema_transformer3d.to(accelerator.device) + + # Move text_encode and vae to gpu and cast to weight_dtype + vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device if not args.low_vram else "cpu") + clip_image_encoder.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype) + audio_encoder.to(accelerator.device if not args.low_vram else "cpu", dtype=torch.float32) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + tracker_config = dict(vars(args)) + keys_to_pop = [k for k, v in tracker_config.items() if isinstance(v, list)] + for k in keys_to_pop: + tracker_config.pop(k) + print(f"Removed tracker_config['{k}']") + accelerator.init_trackers(args.tracker_project_name, tracker_config) + + # Function for unwrapping if model was compiled with `torch.compile`. + def unwrap_model(model): + model = accelerator.unwrap_model(model) + model = model._orig_mod if is_compiled_module(model) else model + return model + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + initial_global_step = 0 + else: + global_step = int(path.split("-")[1]) + + initial_global_step = global_step + + pkl_path = os.path.join(os.path.join(args.output_dir, path), "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + _, first_epoch = pickle.load(file) + else: + first_epoch = global_step // num_update_steps_per_epoch + print(f"Load pkl from {pkl_path}. Get first_epoch = {first_epoch}.") + + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + else: + initial_global_step = 0 + + progress_bar = tqdm( + range(0, args.max_train_steps), + initial=initial_global_step, + desc="Steps", + # Only show the progress bar once on each machine. + disable=not accelerator.is_local_main_process, + ) + + if args.multi_stream: + # create extra cuda streams to speedup inpaint vae computation + vae_stream_1 = torch.cuda.Stream() + vae_stream_2 = torch.cuda.Stream() + else: + vae_stream_1 = None + vae_stream_2 = None + + idx_sampling = DiscreteSampling(args.train_sampling_steps, uniform_sampling=args.uniform_sampling) + + for epoch in range(first_epoch, args.num_train_epochs): + train_loss = 0.0 + batch_sampler.sampler.generator = torch.Generator().manual_seed(args.seed + epoch) + for step, batch in enumerate(train_dataloader): + # Data batch sanity check + if epoch == first_epoch and step == 0: + pixel_values, texts = batch['pixel_values'].cpu(), batch['text'] + pixel_values = rearrange(pixel_values, "b f c h w -> b c f h w") + os.makedirs(os.path.join(args.output_dir, "sanity_check"), exist_ok=True) + for idx, (pixel_value, text) in enumerate(zip(pixel_values, texts)): + pixel_value = pixel_value[None, ...] + gif_name = '-'.join(text.replace('/', '').split()[:10]) if not text == '' else f'{global_step}-{idx}' + save_videos_grid(pixel_value, f"{args.output_dir}/sanity_check/{gif_name[:10]}.gif", rescale=True) + clip_pixel_values, mask_pixel_values, texts = batch['clip_pixel_values'].cpu(), batch['mask_pixel_values'].cpu(), batch['text'] + mask_pixel_values = rearrange(mask_pixel_values, "b f c h w -> b c f h w") + for idx, (clip_pixel_value, pixel_value, text) in enumerate(zip(clip_pixel_values, mask_pixel_values, texts)): + pixel_value = pixel_value[None, ...] + Image.fromarray(np.uint8(clip_pixel_value)).save(f"{args.output_dir}/sanity_check/clip_{gif_name[:10] if not text == '' else f'{global_step}-{idx}'}.png") + save_videos_grid(pixel_value, f"{args.output_dir}/sanity_check/mask_{gif_name[:10] if not text == '' else f'{global_step}-{idx}'}.gif", rescale=True) + + with accelerator.accumulate(transformer3d): + # Convert images to latent space + pixel_values = batch["pixel_values"].to(weight_dtype) + audio = batch["audio"] + sample_rate = batch["sample_rate"] + + # Increase the batch size when the length of the latent sequence of the current sample is small + if args.auto_tile_batch_size and args.training_with_video_token_length and zero_stage != 3: + if args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 16 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + pixel_values = torch.tile(pixel_values, (4, 1, 1, 1, 1)) + if args.enable_text_encoder_in_dataloader: + batch['encoder_hidden_states'] = torch.tile(batch['encoder_hidden_states'], (4, 1, 1)) + batch['encoder_attention_mask'] = torch.tile(batch['encoder_attention_mask'], (4, 1)) + else: + batch['text'] = batch['text'] * 4 + elif args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 4 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + pixel_values = torch.tile(pixel_values, (2, 1, 1, 1, 1)) + if args.enable_text_encoder_in_dataloader: + batch['encoder_hidden_states'] = torch.tile(batch['encoder_hidden_states'], (2, 1, 1)) + batch['encoder_attention_mask'] = torch.tile(batch['encoder_attention_mask'], (2, 1)) + else: + batch['text'] = batch['text'] * 2 + + clip_pixel_values = batch["clip_pixel_values"].to(weight_dtype) + mask_pixel_values = batch["mask_pixel_values"].to(weight_dtype) + mask = batch["mask"].to(weight_dtype) + # Increase the batch size when the length of the latent sequence of the current sample is small + if args.auto_tile_batch_size and args.training_with_video_token_length and zero_stage != 3: + if args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 16 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + clip_pixel_values = torch.tile(clip_pixel_values, (4, 1, 1, 1)) + mask_pixel_values = torch.tile(mask_pixel_values, (4, 1, 1, 1, 1)) + mask = torch.tile(mask, (4, 1, 1, 1, 1)) + elif args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 4 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + clip_pixel_values = torch.tile(clip_pixel_values, (2, 1, 1, 1)) + mask_pixel_values = torch.tile(mask_pixel_values, (2, 1, 1, 1, 1)) + mask = torch.tile(mask, (2, 1, 1, 1, 1)) + + if args.random_frame_crop: + def _create_special_list(length): + if length == 1: + return [1.0] + if length >= 2: + last_element = 0.90 + remaining_sum = 1.0 - last_element + other_elements_value = remaining_sum / (length - 1) + special_list = [other_elements_value] * (length - 1) + [last_element] + return special_list + select_frames = [_tmp for _tmp in list(range(sample_n_frames_bucket_interval + 1, args.video_sample_n_frames + sample_n_frames_bucket_interval, sample_n_frames_bucket_interval))] + select_frames_prob = np.array(_create_special_list(len(select_frames))) + + if len(select_frames) != 0: + if rng is None: + temp_n_frames = np.random.choice(select_frames, p = select_frames_prob) + else: + temp_n_frames = rng.choice(select_frames, p = select_frames_prob) + else: + temp_n_frames = 1 + + # Magvae needs the number of frames to be 4n + 1. + temp_n_frames = (temp_n_frames - 1) // sample_n_frames_bucket_interval + 1 + + pixel_values = pixel_values[:, :temp_n_frames, :, :] + + mask_pixel_values = mask_pixel_values[:, :temp_n_frames, :, :] + mask = mask[:, :temp_n_frames, :, :] + + # Keep all node same token length to accelerate the traning when resolution grows. + if args.keep_all_node_same_token_length: + if args.token_sample_size > 256: + numbers_list = list(range(256, args.token_sample_size + 1, 128)) + + if numbers_list[-1] != args.token_sample_size: + numbers_list.append(args.token_sample_size) + else: + numbers_list = [256] + numbers_list = [_number * _number * args.video_sample_n_frames for _number in numbers_list] + + actual_token_length = index_rng.choice(numbers_list) + actual_video_length = (min( + actual_token_length / pixel_values.size()[-1] / pixel_values.size()[-2], args.video_sample_n_frames + ) - 1) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1 + actual_video_length = int(max(actual_video_length, 1)) + + # Magvae needs the number of frames to be 4n + 1. + actual_video_length = (actual_video_length - 1) // sample_n_frames_bucket_interval + 1 + + pixel_values = pixel_values[:, :actual_video_length, :, :] + mask_pixel_values = mask_pixel_values[:, :actual_video_length, :, :] + mask = mask[:, :actual_video_length, :, :] + + # Make the inpaint latents to be zeros. + t2v_flag = [(_mask == 1).all() for _mask in mask] + new_t2v_flag = [] + for _mask in t2v_flag: + if _mask and np.random.rand() < 0.90: + new_t2v_flag.append(0) + else: + new_t2v_flag.append(1) + t2v_flag = torch.from_numpy(np.array(new_t2v_flag)).to(accelerator.device, dtype=weight_dtype) + + if args.low_vram: + torch.cuda.empty_cache() + vae.to(accelerator.device) + clip_image_encoder.to(accelerator.device) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to("cpu") + + with torch.no_grad(): + # This way is quicker when batch grows up + def _batch_encode_vae(pixel_values): + pixel_values = rearrange(pixel_values, "b f c h w -> b c f h w") + bs = args.vae_mini_batch + new_pixel_values = [] + for i in range(0, pixel_values.shape[0], bs): + pixel_values_bs = pixel_values[i : i + bs] + pixel_values_bs = vae.encode(pixel_values_bs)[0] + pixel_values_bs = pixel_values_bs.sample() + new_pixel_values.append(pixel_values_bs) + return torch.cat(new_pixel_values, dim = 0) + if vae_stream_1 is not None: + vae_stream_1.wait_stream(torch.cuda.current_stream()) + with torch.cuda.stream(vae_stream_1): + latents = _batch_encode_vae(pixel_values) + else: + latents = _batch_encode_vae(pixel_values) + + mask = rearrange(mask, "b f c h w -> b c f h w") + mask = torch.concat( + [ + torch.repeat_interleave(mask[:, :, 0:1], repeats=4, dim=2), + mask[:, :, 1:] + ], dim=2 + ) + mask = mask.view(mask.shape[0], mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]) + mask = mask.transpose(1, 2) + mask = resize_mask(1 - mask, latents) + + # Encode inpaint latents. + mask_latents = _batch_encode_vae(mask_pixel_values) + if vae_stream_2 is not None: + torch.cuda.current_stream().wait_stream(vae_stream_2) + + inpaint_latents = torch.concat([mask, mask_latents], dim=1) + inpaint_latents = t2v_flag[:, None, None, None, None] * inpaint_latents + + clip_context = [] + for clip_pixel_value in clip_pixel_values: + clip_image = Image.fromarray(np.uint8(clip_pixel_value.float().cpu().numpy())) + clip_image = TF.to_tensor(clip_image).sub_(0.5).div_(0.5).to(clip_image_encoder.device, weight_dtype) + _clip_context = clip_image_encoder([clip_image[:, None, :, :]]) + + if rng is None: + zero_init_clip_in = np.random.choice([True, False], p=[0.1, 0.9]) + else: + zero_init_clip_in = rng.choice([True, False], p=[0.1, 0.9]) + clip_context.append(_clip_context if not zero_init_clip_in else torch.zeros_like(_clip_context)) + + clip_context = torch.cat(clip_context) + + # wait for latents = vae.encode(pixel_values) to complete + if vae_stream_1 is not None: + torch.cuda.current_stream().wait_stream(vae_stream_1) + + if args.low_vram: + vae.to('cpu') + clip_image_encoder.to('cpu') + torch.cuda.empty_cache() + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device) + + if args.enable_text_encoder_in_dataloader: + prompt_embeds = batch['encoder_hidden_states'].to(device=latents.device) + else: + with torch.no_grad(): + prompt_ids = tokenizer( + batch['text'], + padding="max_length", + max_length=args.tokenizer_max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt" + ) + text_input_ids = prompt_ids.input_ids + prompt_attention_mask = prompt_ids.attention_mask + + seq_lens = prompt_attention_mask.gt(0).sum(dim=1).long() + prompt_embeds = text_encoder(text_input_ids.to(latents.device), attention_mask=prompt_attention_mask.to(latents.device))[0] + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + + with torch.no_grad(): + # Extract audio emb + audio_wav2vec_fea = [] + for index in range(len(audio)): + _audio_wav2vec_fea = audio_encoder.extract_audio_feat_without_file_load(audio[index], sample_rate[index]) + audio_wav2vec_fea.append(_audio_wav2vec_fea) + audio_wav2vec_fea = torch.cat(audio_wav2vec_fea).to(weight_dtype) + + if args.low_vram and not args.enable_text_encoder_in_dataloader: + text_encoder.to('cpu') + torch.cuda.empty_cache() + + bsz, channel, num_frames, height, width = latents.size() + noise = torch.randn(latents.size(), device=latents.device, generator=torch_rng, dtype=weight_dtype) + + if not args.uniform_sampling: + u = compute_density_for_timestep_sampling( + weighting_scheme=args.weighting_scheme, + batch_size=bsz, + logit_mean=args.logit_mean, + logit_std=args.logit_std, + mode_scale=args.mode_scale, + ) + indices = (u * noise_scheduler.config.num_train_timesteps).long() + else: + # Sample a random timestep for each image + # timesteps = generate_timestep_with_lognorm(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + # timesteps = torch.randint(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + indices = idx_sampling(bsz, generator=torch_rng, device=latents.device) + indices = indices.long().cpu() + timesteps = noise_scheduler.timesteps[indices].to(device=latents.device) + + def get_sigmas(timesteps, n_dim=4, dtype=torch.float32): + sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype) + schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device) + timesteps = timesteps.to(accelerator.device) + step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < n_dim: + sigma = sigma.unsqueeze(-1) + return sigma + + # Add noise according to flow matching. + # zt = (1 - texp) * x + texp * z1 + sigmas = get_sigmas(timesteps, n_dim=latents.ndim, dtype=latents.dtype) + noisy_latents = (1.0 - sigmas) * latents + sigmas * noise + + # Add noise + target = noise - latents + + target_shape = (vae.latent_channels, num_frames, width, height) + seq_len = math.ceil( + (target_shape[2] * target_shape[3]) / + (accelerator.unwrap_model(transformer3d).config.patch_size[1] * accelerator.unwrap_model(transformer3d).config.patch_size[2]) * + target_shape[1] + ) + + # Predict the noise residual + if rng is None: + audio_scale = np.random.choice([1, 0], p=[0.9, 0.1]) + else: + audio_scale = rng.choice([1, 0], p=[0.9, 0.1]) + with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device): + noise_pred = transformer3d( + x=noisy_latents, + context=prompt_embeds, + t=timesteps, + seq_len=seq_len, + y=inpaint_latents, + clip_fea=clip_context, + audio_wav2vec_fea=audio_wav2vec_fea, + audio_scale=audio_scale, + ) + + def custom_mse_loss(noise_pred, target, weighting=None, threshold=50): + noise_pred = noise_pred.float() + target = target.float() + diff = noise_pred - target + mse_loss = F.mse_loss(noise_pred, target, reduction='none') + mask = (diff.abs() <= threshold).float() + masked_loss = mse_loss * mask + if weighting is not None: + masked_loss = masked_loss * weighting + final_loss = masked_loss.mean() + return final_loss + + weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas) + loss = custom_mse_loss(noise_pred.float(), target.float(), weighting.float()) + loss = loss.mean() + + if args.motion_sub_loss and noise_pred.size()[2] > 2: + gt_sub_noise = noise_pred[:, :, 1:].float() - noise_pred[:, :, :-1].float() + pre_sub_noise = target[:, :, 1:].float() - target[:, :, :-1].float() + sub_loss = F.mse_loss(gt_sub_noise, pre_sub_noise, reduction="mean") + loss = loss * (1 - args.motion_sub_loss_ratio) + sub_loss * args.motion_sub_loss_ratio + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + if not args.use_deepspeed and not args.use_fsdp: + trainable_params_grads = [p.grad for p in trainable_params if p.grad is not None] + trainable_params_total_norm = torch.norm(torch.stack([torch.norm(g.detach(), 2) for g in trainable_params_grads]), 2) + max_grad_norm = linear_decay(args.max_grad_norm * args.initial_grad_norm_ratio, args.max_grad_norm, args.abnormal_norm_clip_start, global_step) + if trainable_params_total_norm / max_grad_norm > 5 and global_step > args.abnormal_norm_clip_start: + actual_max_grad_norm = max_grad_norm / min((trainable_params_total_norm / max_grad_norm), 10) + else: + actual_max_grad_norm = max_grad_norm + else: + actual_max_grad_norm = args.max_grad_norm + + if not args.use_deepspeed and not args.use_fsdp and args.report_model_info and accelerator.is_main_process: + if trainable_params_total_norm > 1 and global_step > args.abnormal_norm_clip_start: + for name, param in transformer3d.named_parameters(): + if param.requires_grad: + writer.add_scalar(f'gradients/before_clip_norm/{name}', param.grad.norm(), global_step=global_step) + + norm_sum = accelerator.clip_grad_norm_(trainable_params, actual_max_grad_norm) + if not args.use_deepspeed and not args.use_fsdp and args.report_model_info and accelerator.is_main_process: + writer.add_scalar(f'gradients/norm_sum', norm_sum, global_step=global_step) + writer.add_scalar(f'gradients/actual_max_grad_norm', actual_max_grad_norm, global_step=global_step) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + + if args.use_ema: + ema_transformer3d.step(transformer3d.parameters()) + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if global_step % args.checkpointing_steps == 0: + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + if accelerator.is_main_process: + if args.validation_prompts is not None and global_step % args.validation_steps == 0: + if args.use_ema: + # Store the UNet parameters temporarily and load the EMA parameters to perform inference. + ema_transformer3d.store(transformer3d.parameters()) + ema_transformer3d.copy_to(transformer3d.parameters()) + log_validation( + vae, + text_encoder, + tokenizer, + clip_image_encoder, + transformer3d, + args, + config, + accelerator, + weight_dtype, + global_step, + ) + if args.use_ema: + # Switch back to the original transformer3d parameters. + ema_transformer3d.restore(transformer3d.parameters()) + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompts is not None and epoch % args.validation_epochs == 0: + if args.use_ema: + # Store the UNet parameters temporarily and load the EMA parameters to perform inference. + ema_transformer3d.store(transformer3d.parameters()) + ema_transformer3d.copy_to(transformer3d.parameters()) + log_validation( + vae, + text_encoder, + tokenizer, + clip_image_encoder, + transformer3d, + args, + config, + accelerator, + weight_dtype, + global_step, + ) + if args.use_ema: + # Switch back to the original transformer3d parameters. + ema_transformer3d.restore(transformer3d.parameters()) + + # Create the pipeline using the trained modules and save it. + accelerator.wait_for_everyone() + if accelerator.is_main_process: + transformer3d = unwrap_model(transformer3d) + if args.use_ema: + ema_transformer3d.copy_to(transformer3d.parameters()) + + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + accelerator.end_training() + + +if __name__ == "__main__": + main() diff --git a/scripts/fantasytalking/train.sh b/scripts/fantasytalking/train.sh new file mode 100644 index 0000000..225a556 --- /dev/null +++ b/scripts/fantasytalking/train.sh @@ -0,0 +1,40 @@ +export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-I2V-14B-720P" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" scripts/fantasytalking/train.py \ + --config_path="config/wan2.1/wan_civitai.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --video_sample_size=512 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=2e-05 \ + --lr_scheduler="constant_with_warmup" \ + --lr_warmup_steps=100 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --low_vram \ + --transformer_path="models/FantasyTalking/fantasytalking_model.ckpt" \ + --trainable_modules "processor." "proj_model." \ No newline at end of file diff --git a/videox_fun/data/__init__.py b/videox_fun/data/__init__.py new file mode 100644 index 0000000..6ae2811 --- /dev/null +++ b/videox_fun/data/__init__.py @@ -0,0 +1,9 @@ +from .dataset_image import CC15M, ImageEditDataset +from .dataset_image_video import (ImageVideoControlDataset, ImageVideoDataset, + ImageVideoSampler) +from .dataset_video import VideoDataset, VideoSpeechDataset, WebVid10M +from .utils import (VIDEO_READER_TIMEOUT, Camera, VideoReader_contextmanager, + custom_meshgrid, get_random_mask, get_relative_pose, + get_video_reader_batch, padding_image, process_pose_file, + process_pose_params, ray_condition, resize_frame, + resize_image_with_target_area) diff --git a/videox_fun/data/dataset_image.py b/videox_fun/data/dataset_image.py index f416126..18c672a 100644 --- a/videox_fun/data/dataset_image.py +++ b/videox_fun/data/dataset_image.py @@ -182,7 +182,7 @@ class ImageEditDataset(Dataset): if __name__ == "__main__": dataset = CC15M( - csv_path="/mnt_wg/zhoumo.xjq/CCUtils/cc15m_add_index.json", + csv_path="./cc15m_add_index.json", resolution=512, ) diff --git a/videox_fun/data/dataset_image_video.py b/videox_fun/data/dataset_image_video.py index 23940d8..6d2252f 100755 --- a/videox_fun/data/dataset_image_video.py +++ b/videox_fun/data/dataset_image_video.py @@ -24,238 +24,12 @@ from safetensors.torch import load_file from torch.utils.data import BatchSampler, Sampler from torch.utils.data.dataset import Dataset -VIDEO_READER_TIMEOUT = 20 +from .utils import (VIDEO_READER_TIMEOUT, Camera, VideoReader_contextmanager, + custom_meshgrid, get_random_mask, get_relative_pose, + get_video_reader_batch, padding_image, process_pose_file, + process_pose_params, ray_condition, resize_frame, + resize_image_with_target_area) -def get_random_mask(shape, image_start_only=False): - f, c, h, w = shape - mask = torch.zeros((f, 1, h, w), dtype=torch.uint8) - - if not image_start_only: - if f != 1: - mask_index = np.random.choice([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], p=[0.05, 0.2, 0.2, 0.2, 0.05, 0.05, 0.05, 0.1, 0.05, 0.05]) - else: - mask_index = np.random.choice([0, 1], p = [0.2, 0.8]) - if mask_index == 0: - center_x = torch.randint(0, w, (1,)).item() - center_y = torch.randint(0, h, (1,)).item() - block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item() # 方块的宽度范围 - block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item() # 方块的高度范围 - - start_x = max(center_x - block_size_x // 2, 0) - end_x = min(center_x + block_size_x // 2, w) - start_y = max(center_y - block_size_y // 2, 0) - end_y = min(center_y + block_size_y // 2, h) - mask[:, :, start_y:end_y, start_x:end_x] = 1 - elif mask_index == 1: - mask[:, :, :, :] = 1 - elif mask_index == 2: - mask_frame_index = np.random.randint(1, 5) - mask[mask_frame_index:, :, :, :] = 1 - elif mask_index == 3: - mask_frame_index = np.random.randint(1, 5) - mask[mask_frame_index:-mask_frame_index, :, :, :] = 1 - elif mask_index == 4: - center_x = torch.randint(0, w, (1,)).item() - center_y = torch.randint(0, h, (1,)).item() - block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item() # 方块的宽度范围 - block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item() # 方块的高度范围 - - start_x = max(center_x - block_size_x // 2, 0) - end_x = min(center_x + block_size_x // 2, w) - start_y = max(center_y - block_size_y // 2, 0) - end_y = min(center_y + block_size_y // 2, h) - - mask_frame_before = np.random.randint(0, f // 2) - mask_frame_after = np.random.randint(f // 2, f) - mask[mask_frame_before:mask_frame_after, :, start_y:end_y, start_x:end_x] = 1 - elif mask_index == 5: - mask = torch.randint(0, 2, (f, 1, h, w), dtype=torch.uint8) - elif mask_index == 6: - num_frames_to_mask = random.randint(1, max(f // 2, 1)) - frames_to_mask = random.sample(range(f), num_frames_to_mask) - - for i in frames_to_mask: - block_height = random.randint(1, h // 4) - block_width = random.randint(1, w // 4) - top_left_y = random.randint(0, h - block_height) - top_left_x = random.randint(0, w - block_width) - mask[i, 0, top_left_y:top_left_y + block_height, top_left_x:top_left_x + block_width] = 1 - elif mask_index == 7: - center_x = torch.randint(0, w, (1,)).item() - center_y = torch.randint(0, h, (1,)).item() - a = torch.randint(min(w, h) // 8, min(w, h) // 4, (1,)).item() # 长半轴 - b = torch.randint(min(h, w) // 8, min(h, w) // 4, (1,)).item() # 短半轴 - - for i in range(h): - for j in range(w): - if ((i - center_y) ** 2) / (b ** 2) + ((j - center_x) ** 2) / (a ** 2) < 1: - mask[:, :, i, j] = 1 - elif mask_index == 8: - center_x = torch.randint(0, w, (1,)).item() - center_y = torch.randint(0, h, (1,)).item() - radius = torch.randint(min(h, w) // 8, min(h, w) // 4, (1,)).item() - for i in range(h): - for j in range(w): - if (i - center_y) ** 2 + (j - center_x) ** 2 < radius ** 2: - mask[:, :, i, j] = 1 - elif mask_index == 9: - for idx in range(f): - if np.random.rand() > 0.5: - mask[idx, :, :, :] = 1 - else: - raise ValueError(f"The mask_index {mask_index} is not define") - else: - if f != 1: - mask[1:, :, :, :] = 1 - else: - mask[:, :, :, :] = 1 - return mask - -class Camera(object): - """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py - """ - def __init__(self, entry): - fx, fy, cx, cy = entry[1:5] - self.fx = fx - self.fy = fy - self.cx = cx - self.cy = cy - w2c_mat = np.array(entry[7:]).reshape(3, 4) - w2c_mat_4x4 = np.eye(4) - w2c_mat_4x4[:3, :] = w2c_mat - self.w2c_mat = w2c_mat_4x4 - self.c2w_mat = np.linalg.inv(w2c_mat_4x4) - -def custom_meshgrid(*args): - """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py - """ - # ref: https://pytorch.org/docs/stable/generated/torch.meshgrid.html?highlight=meshgrid#torch.meshgrid - if pver.parse(torch.__version__) < pver.parse('1.10'): - return torch.meshgrid(*args) - else: - return torch.meshgrid(*args, indexing='ij') - -def get_relative_pose(cam_params): - """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py - """ - abs_w2cs = [cam_param.w2c_mat for cam_param in cam_params] - abs_c2ws = [cam_param.c2w_mat for cam_param in cam_params] - cam_to_origin = 0 - target_cam_c2w = np.array([ - [1, 0, 0, 0], - [0, 1, 0, -cam_to_origin], - [0, 0, 1, 0], - [0, 0, 0, 1] - ]) - abs2rel = target_cam_c2w @ abs_w2cs[0] - ret_poses = [target_cam_c2w, ] + [abs2rel @ abs_c2w for abs_c2w in abs_c2ws[1:]] - ret_poses = np.array(ret_poses, dtype=np.float32) - return ret_poses - -def ray_condition(K, c2w, H, W, device): - """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py - """ - # c2w: B, V, 4, 4 - # K: B, V, 4 - - B = K.shape[0] - - j, i = custom_meshgrid( - torch.linspace(0, H - 1, H, device=device, dtype=c2w.dtype), - torch.linspace(0, W - 1, W, device=device, dtype=c2w.dtype), - ) - i = i.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5 # [B, HxW] - j = j.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5 # [B, HxW] - - fx, fy, cx, cy = K.chunk(4, dim=-1) # B,V, 1 - - zs = torch.ones_like(i) # [B, HxW] - xs = (i - cx) / fx * zs - ys = (j - cy) / fy * zs - zs = zs.expand_as(ys) - - directions = torch.stack((xs, ys, zs), dim=-1) # B, V, HW, 3 - directions = directions / directions.norm(dim=-1, keepdim=True) # B, V, HW, 3 - - rays_d = directions @ c2w[..., :3, :3].transpose(-1, -2) # B, V, 3, HW - rays_o = c2w[..., :3, 3] # B, V, 3 - rays_o = rays_o[:, :, None].expand_as(rays_d) # B, V, 3, HW - # c2w @ dirctions - rays_dxo = torch.cross(rays_o, rays_d) - plucker = torch.cat([rays_dxo, rays_d], dim=-1) - plucker = plucker.reshape(B, c2w.shape[1], H, W, 6) # B, V, H, W, 6 - # plucker = plucker.permute(0, 1, 4, 2, 3) - return plucker - -def process_pose_file(pose_file_path, width=672, height=384, original_pose_width=1280, original_pose_height=720, device='cpu', return_poses=False): - """Modified from https://github.com/hehao13/CameraCtrl/blob/main/inference.py - """ - with open(pose_file_path, 'r') as f: - poses = f.readlines() - - poses = [pose.strip().split(' ') for pose in poses[1:]] - cam_params = [[float(x) for x in pose] for pose in poses] - if return_poses: - return cam_params - else: - cam_params = [Camera(cam_param) for cam_param in cam_params] - - sample_wh_ratio = width / height - pose_wh_ratio = original_pose_width / original_pose_height # Assuming placeholder ratios, change as needed - - if pose_wh_ratio > sample_wh_ratio: - resized_ori_w = height * pose_wh_ratio - for cam_param in cam_params: - cam_param.fx = resized_ori_w * cam_param.fx / width - else: - resized_ori_h = width / pose_wh_ratio - for cam_param in cam_params: - cam_param.fy = resized_ori_h * cam_param.fy / height - - intrinsic = np.asarray([[cam_param.fx * width, - cam_param.fy * height, - cam_param.cx * width, - cam_param.cy * height] - for cam_param in cam_params], dtype=np.float32) - - K = torch.as_tensor(intrinsic)[None] # [1, 1, 4] - c2ws = get_relative_pose(cam_params) # Assuming this function is defined elsewhere - c2ws = torch.as_tensor(c2ws)[None] # [1, n_frame, 4, 4] - plucker_embedding = ray_condition(K, c2ws, height, width, device=device)[0].permute(0, 3, 1, 2).contiguous() # V, 6, H, W - plucker_embedding = plucker_embedding[None] - plucker_embedding = rearrange(plucker_embedding, "b f c h w -> b f h w c")[0] - return plucker_embedding - -def process_pose_params(cam_params, width=672, height=384, original_pose_width=1280, original_pose_height=720, device='cpu'): - """Modified from https://github.com/hehao13/CameraCtrl/blob/main/inference.py - """ - cam_params = [Camera(cam_param) for cam_param in cam_params] - - sample_wh_ratio = width / height - pose_wh_ratio = original_pose_width / original_pose_height # Assuming placeholder ratios, change as needed - - if pose_wh_ratio > sample_wh_ratio: - resized_ori_w = height * pose_wh_ratio - for cam_param in cam_params: - cam_param.fx = resized_ori_w * cam_param.fx / width - else: - resized_ori_h = width / pose_wh_ratio - for cam_param in cam_params: - cam_param.fy = resized_ori_h * cam_param.fy / height - - intrinsic = np.asarray([[cam_param.fx * width, - cam_param.fy * height, - cam_param.cx * width, - cam_param.cy * height] - for cam_param in cam_params], dtype=np.float32) - - K = torch.as_tensor(intrinsic)[None] # [1, 1, 4] - c2ws = get_relative_pose(cam_params) # Assuming this function is defined elsewhere - c2ws = torch.as_tensor(c2ws)[None] # [1, n_frame, 4, 4] - plucker_embedding = ray_condition(K, c2ws, height, width, device=device)[0].permute(0, 3, 1, 2).contiguous() # V, 6, H, W - plucker_embedding = plucker_embedding[None] - plucker_embedding = rearrange(plucker_embedding, "b f c h w -> b f h w c")[0] - return plucker_embedding class ImageVideoSampler(BatchSampler): """A sampler wrapper for grouping images with similar aspect ratio into a same batch. @@ -304,34 +78,6 @@ class ImageVideoSampler(BatchSampler): yield bucket[:] del bucket[:] -@contextmanager -def VideoReader_contextmanager(*args, **kwargs): - vr = VideoReader(*args, **kwargs) - try: - yield vr - finally: - del vr - gc.collect() - -def get_video_reader_batch(video_reader, batch_index): - frames = video_reader.get_batch(batch_index).asnumpy() - return frames - -def resize_frame(frame, target_short_side): - h, w, _ = frame.shape - if h < w: - if target_short_side > h: - return frame - new_h = target_short_side - new_w = int(target_short_side * w / h) - else: - if target_short_side > w: - return frame - new_w = target_short_side - new_h = int(target_short_side * h / w) - - resized_frame = cv2.resize(frame, (new_w, new_h)) - return resized_frame class ImageVideoDataset(Dataset): def __init__( @@ -513,33 +259,6 @@ class ImageVideoDataset(Dataset): return sample -def padding_image(images, new_width, new_height): - new_image = Image.new('RGB', (new_width, new_height), (255, 255, 255)) - - aspect_ratio = images.width / images.height - if new_width / new_height > 1: - if aspect_ratio > new_width / new_height: - new_img_width = new_width - new_img_height = int(new_img_width / aspect_ratio) - else: - new_img_height = new_height - new_img_width = int(new_img_height * aspect_ratio) - else: - if aspect_ratio > new_width / new_height: - new_img_width = new_width - new_img_height = int(new_img_width / aspect_ratio) - else: - new_img_height = new_height - new_img_width = int(new_img_height * aspect_ratio) - - resized_img = images.resize((new_img_width, new_img_height)) - - paste_x = (new_width - new_img_width) // 2 - paste_y = (new_height - new_img_height) // 2 - - new_image.paste(resized_img, (paste_x, paste_y)) - - return new_image class ImageVideoControlDataset(Dataset): def __init__( @@ -556,6 +275,7 @@ class ImageVideoControlDataset(Dataset): enable_camera_info=False, return_file_name=False, enable_subject_info=False, + padding_subject_info=True, ): # Loading annotations from files print(f"loading annotations from {ann_path} ...") @@ -590,6 +310,7 @@ class ImageVideoControlDataset(Dataset): self.enable_inpaint = enable_inpaint self.enable_camera_info = enable_camera_info self.enable_subject_info = enable_subject_info + self.padding_subject_info = padding_subject_info self.video_length_drop_start = video_length_drop_start self.video_length_drop_end = video_length_drop_end @@ -757,11 +478,18 @@ class ImageVideoControlDataset(Dataset): width, height = subject_image.size total_pixels = width * height - img = padding_image(subject_image, visual_width, visual_height) + if self.padding_subject_info: + img = padding_image(subject_image, visual_width, visual_height) + else: + img = resize_image_with_target_area(subject_image, 1024 * 1024) + if random.random() < 0.5: img = img.transpose(Image.FLIP_LEFT_RIGHT) - subject_images.append(img) - subject_image = np.array(subject_images) + subject_images.append(np.array(img)) + if self.padding_subject_info: + subject_image = np.array(subject_images) + else: + subject_image = subject_images else: subject_image = None @@ -806,15 +534,23 @@ class ImageVideoControlDataset(Dataset): width, height = subject_image.size total_pixels = width * height - img = padding_image(subject_image, visual_width, visual_height) + if self.padding_subject_info: + img = padding_image(subject_image, visual_width, visual_height) + else: + img = resize_image_with_target_area(subject_image, 1024 * 1024) + if random.random() < 0.5: img = img.transpose(Image.FLIP_LEFT_RIGHT) - subject_images.append(img) - subject_image = np.array(subject_images) + subject_images.append(np.array(img)) + if self.padding_subject_info: + subject_image = np.array(subject_images) + else: + subject_image = subject_images else: subject_image = None return image, control_image, subject_image, None, text, 'image' + def __len__(self): return self.length diff --git a/videox_fun/data/dataset_video.py b/videox_fun/data/dataset_video.py index c78367d..84ad7a4 100644 --- a/videox_fun/data/dataset_video.py +++ b/videox_fun/data/dataset_video.py @@ -10,6 +10,7 @@ from threading import Thread import albumentations import cv2 +import librosa import numpy as np import torch import torchvision.transforms as transforms @@ -20,61 +21,11 @@ from PIL import Image from torch.utils.data import BatchSampler, Sampler from torch.utils.data.dataset import Dataset -VIDEO_READER_TIMEOUT = 20 - -def get_random_mask(shape): - f, c, h, w = shape - - mask_index = np.random.randint(0, 4) - mask = torch.zeros((f, 1, h, w), dtype=torch.uint8) - if mask_index == 0: - mask[1:, :, :, :] = 1 - elif mask_index == 1: - mask_frame_index = 1 - mask[mask_frame_index:-mask_frame_index, :, :, :] = 1 - elif mask_index == 2: - center_x = torch.randint(0, w, (1,)).item() - center_y = torch.randint(0, h, (1,)).item() - block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item() # 方块的宽度范围 - block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item() # 方块的高度范围 - - start_x = max(center_x - block_size_x // 2, 0) - end_x = min(center_x + block_size_x // 2, w) - start_y = max(center_y - block_size_y // 2, 0) - end_y = min(center_y + block_size_y // 2, h) - mask[:, :, start_y:end_y, start_x:end_x] = 1 - elif mask_index == 3: - center_x = torch.randint(0, w, (1,)).item() - center_y = torch.randint(0, h, (1,)).item() - block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item() # 方块的宽度范围 - block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item() # 方块的高度范围 - - start_x = max(center_x - block_size_x // 2, 0) - end_x = min(center_x + block_size_x // 2, w) - start_y = max(center_y - block_size_y // 2, 0) - end_y = min(center_y + block_size_y // 2, h) - - mask_frame_before = np.random.randint(0, f // 2) - mask_frame_after = np.random.randint(f // 2, f) - mask[mask_frame_before:mask_frame_after, :, start_y:end_y, start_x:end_x] = 1 - else: - raise ValueError(f"The mask_index {mask_index} is not define") - return mask - - -@contextmanager -def VideoReader_contextmanager(*args, **kwargs): - vr = VideoReader(*args, **kwargs) - try: - yield vr - finally: - del vr - gc.collect() - - -def get_video_reader_batch(video_reader, batch_index): - frames = video_reader.get_batch(batch_index).asnumpy() - return frames +from .utils import (VIDEO_READER_TIMEOUT, Camera, VideoReader_contextmanager, + custom_meshgrid, get_random_mask, get_relative_pose, + get_video_reader_batch, padding_image, process_pose_file, + process_pose_params, ray_condition, resize_frame, + resize_image_with_target_area) class WebVid10M(Dataset): @@ -157,16 +108,16 @@ class WebVid10M(Dataset): class VideoDataset(Dataset): def __init__( self, - json_path, video_folder=None, + ann_path, data_root=None, sample_size=256, sample_stride=4, sample_n_frames=16, enable_bucket=False, enable_inpaint=False ): - print(f"loading annotations from {json_path} ...") - self.dataset = json.load(open(json_path, 'r')) + print(f"loading annotations from {ann_path} ...") + self.dataset = json.load(open(ann_path, 'r')) self.length = len(self.dataset) print(f"data scale: {self.length}") - self.video_folder = video_folder + self.data_root = data_root self.sample_stride = sample_stride self.sample_n_frames = sample_n_frames self.enable_bucket = enable_bucket @@ -183,19 +134,25 @@ class VideoDataset(Dataset): def get_batch(self, idx): video_dict = self.dataset[idx] - video_id, name = video_dict['file_path'], video_dict['text'] + video_id, text = video_dict['file_path'], video_dict['text'] - if self.video_folder is None: + if self.data_root is None: video_dir = video_id else: - video_dir = os.path.join(self.video_folder, video_id) + video_dir = os.path.join(self.data_root, video_id) with VideoReader_contextmanager(video_dir, num_threads=2) as video_reader: - video_length = len(video_reader) - - clip_length = min(video_length, (self.sample_n_frames - 1) * self.sample_stride + 1) - start_idx = random.randint(0, video_length - clip_length) - batch_index = np.linspace(start_idx, start_idx + clip_length - 1, self.sample_n_frames, dtype=int) + min_sample_n_frames = min( + self.video_sample_n_frames, + int(len(video_reader) * (self.video_length_drop_end - self.video_length_drop_start) // self.video_sample_stride) + ) + if min_sample_n_frames == 0: + raise ValueError(f"No Frames in video.") + + video_length = int(self.video_length_drop_end * len(video_reader)) + clip_length = min(video_length, (min_sample_n_frames - 1) * self.video_sample_stride + 1) + start_idx = random.randint(int(self.video_length_drop_start * video_length), video_length - clip_length) if video_length != clip_length else 0 + batch_index = np.linspace(start_idx, start_idx + clip_length - 1, min_sample_n_frames, dtype=int) try: sample_args = (video_reader, batch_index) @@ -214,29 +171,184 @@ class VideoDataset(Dataset): else: pixel_values = pixel_values - return pixel_values, name + if not self.enable_bucket: + pixel_values = self.video_transforms(pixel_values) + + # Random use no text generation + if random.random() < self.text_drop_ratio: + text = '' + return pixel_values, text def __len__(self): return self.length def __getitem__(self, idx): while True: + sample = {} try: pixel_values, name = self.get_batch(idx) - break + sample["pixel_values"] = pixel_values + sample["text"] = name + sample["idx"] = idx + if len(sample) > 0: + break except Exception as e: - print("Error info:", e) + print(e, self.dataset[idx % len(self.dataset)]) idx = random.randint(0, self.length-1) - if not self.enable_bucket: - pixel_values = self.pixel_transforms(pixel_values) - if self.enable_inpaint: + if self.enable_inpaint and not self.enable_bucket: mask = get_random_mask(pixel_values.size()) - mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * -1 * mask - sample = dict(pixel_values=pixel_values, mask_pixel_values=mask_pixel_values, mask=mask, text=name) + mask_pixel_values = pixel_values * (1 - mask) + torch.zeros_like(pixel_values) * mask + sample["mask_pixel_values"] = mask_pixel_values + sample["mask"] = mask + + clip_pixel_values = sample["pixel_values"][0].permute(1, 2, 0).contiguous() + clip_pixel_values = (clip_pixel_values * 0.5 + 0.5) * 255 + sample["clip_pixel_values"] = clip_pixel_values + + return sample + + +class VideoSpeechDataset(Dataset): + def __init__( + self, + ann_path, data_root=None, + video_sample_size=512, video_sample_stride=4, video_sample_n_frames=16, + enable_bucket=False, enable_inpaint=False, + audio_sr=16000, # 新增：目标音频采样率 + text_drop_ratio=0.1 # 新增：文本丢弃概率 + ): + print(f"loading annotations from {ann_path} ...") + self.dataset = json.load(open(ann_path, 'r')) + self.length = len(self.dataset) + print(f"data scale: {self.length}") + + self.data_root = data_root + self.video_sample_stride = video_sample_stride + self.video_sample_n_frames = video_sample_n_frames + self.enable_bucket = enable_bucket + self.enable_inpaint = enable_inpaint + self.audio_sr = audio_sr + self.text_drop_ratio = text_drop_ratio + + video_sample_size = tuple(video_sample_size) if not isinstance(video_sample_size, int) else (video_sample_size, video_sample_size) + self.pixel_transforms = transforms.Compose( + [ + transforms.Resize(video_sample_size[0]), + transforms.CenterCrop(video_sample_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ] + ) + + def get_batch(self, idx): + video_dict = self.dataset[idx] + video_id, text = video_dict['file_path'], video_dict['text'] + audio_id = video_dict['audio_path'] + + if self.data_root is None: + video_path = video_id else: - sample = dict(pixel_values=pixel_values, text=name) + video_path = os.path.join(self.data_root, video_id) + + if self.data_root is None: + audio_path = audio_id + else: + audio_path = os.path.join(self.data_root, audio_id) + + if not os.path.exists(audio_path): + raise FileNotFoundError(f"Audio file not found for {video_path}") + + with VideoReader_contextmanager(video_path, num_threads=2) as video_reader: + total_frames = len(video_reader) + fps = video_reader.get_avg_fps() # 获取原始视频帧率 + + # 计算实际采样的视频帧数（考虑边界） + max_possible_frames = (total_frames - 1) // self.video_sample_stride + 1 + actual_n_frames = min(self.video_sample_n_frames, max_possible_frames) + if actual_n_frames <= 0: + raise ValueError(f"Video too short: {video_path}") + + # 随机选择起始帧 + max_start = total_frames - (actual_n_frames - 1) * self.video_sample_stride - 1 + start_frame = random.randint(0, max_start) if max_start > 0 else 0 + frame_indices = [start_frame + i * self.video_sample_stride for i in range(actual_n_frames)] + + # 读取视频帧 + try: + sample_args = (video_reader, frame_indices) + pixel_values = func_timeout( + VIDEO_READER_TIMEOUT, get_video_reader_batch, args=sample_args + ) + except FunctionTimedOut: + raise ValueError(f"Read {idx} timeout.") + except Exception as e: + raise ValueError(f"Failed to extract frames from video. Error is {e}.") + + # 视频后处理 + if not self.enable_bucket: + pixel_values = torch.from_numpy(pixel_values).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + pixel_values = self.pixel_transforms(pixel_values) + + # === 新增：加载并截取对应音频 === + # 视频片段的起止时间（秒） + start_time = start_frame / fps + end_time = (start_frame + (actual_n_frames - 1) * self.video_sample_stride) / fps + duration = end_time - start_time + + # 使用 librosa 加载整个音频（或仅加载所需部分，但 librosa.load 不支持精确 seek，所以先加载再切） + audio_input, sample_rate = librosa.load(audio_path, sr=self.audio_sr) # 重采样到目标 sr + + # 转换为样本索引 + start_sample = int(start_time * self.audio_sr) + end_sample = int(end_time * self.audio_sr) + + # 安全截取 + if start_sample >= len(audio_input): + # 音频太短，用零填充或截断 + audio_segment = np.zeros(int(duration * self.audio_sr), dtype=np.float32) + else: + audio_segment = audio_input[start_sample:end_sample] + # 如果太短，补零 + target_len = int(duration * self.audio_sr) + if len(audio_segment) < target_len: + audio_segment = np.pad(audio_segment, (0, target_len - len(audio_segment)), mode='constant') + + # === 文本随机丢弃 === + if random.random() < self.text_drop_ratio: + text = '' + + return pixel_values, text, audio_segment, sample_rate + + def __len__(self): + return self.length + + def __getitem__(self, idx): + while True: + sample = {} + try: + pixel_values, text, audio, sample_rate = self.get_batch(idx) + sample["pixel_values"] = pixel_values + sample["text"] = text + sample["audio"] = torch.from_numpy(audio).float() # 转为 tensor + sample["sample_rate"] = sample_rate + sample["idx"] = idx + break + except Exception as e: + print(f"Error processing {idx}: {e}, retrying with random idx...") + idx = random.randint(0, self.length - 1) + + if self.enable_inpaint and not self.enable_bucket: + mask = get_random_mask(pixel_values.size(), image_start_only=True) + mask_pixel_values = pixel_values * (1 - mask) + torch.zeros_like(pixel_values) * mask + sample["mask_pixel_values"] = mask_pixel_values + sample["mask"] = mask + + clip_pixel_values = sample["pixel_values"][0].permute(1, 2, 0).contiguous() + clip_pixel_values = (clip_pixel_values * 0.5 + 0.5) * 255 + sample["clip_pixel_values"] = clip_pixel_values + return sample diff --git a/videox_fun/data/utils.py b/videox_fun/data/utils.py new file mode 100644 index 0000000..514a41b --- /dev/null +++ b/videox_fun/data/utils.py @@ -0,0 +1,347 @@ +import csv +import gc +import io +import json +import math +import os +import random +from contextlib import contextmanager +from random import shuffle +from threading import Thread + +import albumentations +import cv2 +import numpy as np +import torch +import torch.nn.functional as F +import torchvision.transforms as transforms +from decord import VideoReader +from einops import rearrange +from func_timeout import FunctionTimedOut, func_timeout +from packaging import version as pver +from PIL import Image +from safetensors.torch import load_file +from torch.utils.data import BatchSampler, Sampler +from torch.utils.data.dataset import Dataset + +VIDEO_READER_TIMEOUT = 20 + +def get_random_mask(shape, image_start_only=False): + f, c, h, w = shape + mask = torch.zeros((f, 1, h, w), dtype=torch.uint8) + + if not image_start_only: + if f != 1: + mask_index = np.random.choice([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], p=[0.05, 0.2, 0.2, 0.2, 0.05, 0.05, 0.05, 0.1, 0.05, 0.05]) + else: + mask_index = np.random.choice([0, 1, 7, 8], p = [0.2, 0.7, 0.05, 0.05]) + if mask_index == 0: + center_x = torch.randint(0, w, (1,)).item() + center_y = torch.randint(0, h, (1,)).item() + block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item() # 方块的宽度范围 + block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item() # 方块的高度范围 + + start_x = max(center_x - block_size_x // 2, 0) + end_x = min(center_x + block_size_x // 2, w) + start_y = max(center_y - block_size_y // 2, 0) + end_y = min(center_y + block_size_y // 2, h) + mask[:, :, start_y:end_y, start_x:end_x] = 1 + elif mask_index == 1: + mask[:, :, :, :] = 1 + elif mask_index == 2: + mask_frame_index = np.random.randint(1, 5) + mask[mask_frame_index:, :, :, :] = 1 + elif mask_index == 3: + mask_frame_index = np.random.randint(1, 5) + mask[mask_frame_index:-mask_frame_index, :, :, :] = 1 + elif mask_index == 4: + center_x = torch.randint(0, w, (1,)).item() + center_y = torch.randint(0, h, (1,)).item() + block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item() # 方块的宽度范围 + block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item() # 方块的高度范围 + + start_x = max(center_x - block_size_x // 2, 0) + end_x = min(center_x + block_size_x // 2, w) + start_y = max(center_y - block_size_y // 2, 0) + end_y = min(center_y + block_size_y // 2, h) + + mask_frame_before = np.random.randint(0, f // 2) + mask_frame_after = np.random.randint(f // 2, f) + mask[mask_frame_before:mask_frame_after, :, start_y:end_y, start_x:end_x] = 1 + elif mask_index == 5: + mask = torch.randint(0, 2, (f, 1, h, w), dtype=torch.uint8) + elif mask_index == 6: + num_frames_to_mask = random.randint(1, max(f // 2, 1)) + frames_to_mask = random.sample(range(f), num_frames_to_mask) + + for i in frames_to_mask: + block_height = random.randint(1, h // 4) + block_width = random.randint(1, w // 4) + top_left_y = random.randint(0, h - block_height) + top_left_x = random.randint(0, w - block_width) + mask[i, 0, top_left_y:top_left_y + block_height, top_left_x:top_left_x + block_width] = 1 + elif mask_index == 7: + center_x = torch.randint(0, w, (1,)).item() + center_y = torch.randint(0, h, (1,)).item() + a = torch.randint(min(w, h) // 8, min(w, h) // 4, (1,)).item() # 长半轴 + b = torch.randint(min(h, w) // 8, min(h, w) // 4, (1,)).item() # 短半轴 + + for i in range(h): + for j in range(w): + if ((i - center_y) ** 2) / (b ** 2) + ((j - center_x) ** 2) / (a ** 2) < 1: + mask[:, :, i, j] = 1 + elif mask_index == 8: + center_x = torch.randint(0, w, (1,)).item() + center_y = torch.randint(0, h, (1,)).item() + radius = torch.randint(min(h, w) // 8, min(h, w) // 4, (1,)).item() + for i in range(h): + for j in range(w): + if (i - center_y) ** 2 + (j - center_x) ** 2 < radius ** 2: + mask[:, :, i, j] = 1 + elif mask_index == 9: + for idx in range(f): + if np.random.rand() > 0.5: + mask[idx, :, :, :] = 1 + else: + raise ValueError(f"The mask_index {mask_index} is not define") + else: + if f != 1: + mask[1:, :, :, :] = 1 + else: + mask[:, :, :, :] = 1 + return mask + +@contextmanager +def VideoReader_contextmanager(*args, **kwargs): + vr = VideoReader(*args, **kwargs) + try: + yield vr + finally: + del vr + gc.collect() + +def get_video_reader_batch(video_reader, batch_index): + frames = video_reader.get_batch(batch_index).asnumpy() + return frames + +def resize_frame(frame, target_short_side): + h, w, _ = frame.shape + if h < w: + if target_short_side > h: + return frame + new_h = target_short_side + new_w = int(target_short_side * w / h) + else: + if target_short_side > w: + return frame + new_w = target_short_side + new_h = int(target_short_side * h / w) + + resized_frame = cv2.resize(frame, (new_w, new_h)) + return resized_frame + +def padding_image(images, new_width, new_height): + new_image = Image.new('RGB', (new_width, new_height), (255, 255, 255)) + + aspect_ratio = images.width / images.height + if new_width / new_height > 1: + if aspect_ratio > new_width / new_height: + new_img_width = new_width + new_img_height = int(new_img_width / aspect_ratio) + else: + new_img_height = new_height + new_img_width = int(new_img_height * aspect_ratio) + else: + if aspect_ratio > new_width / new_height: + new_img_width = new_width + new_img_height = int(new_img_width / aspect_ratio) + else: + new_img_height = new_height + new_img_width = int(new_img_height * aspect_ratio) + + resized_img = images.resize((new_img_width, new_img_height)) + + paste_x = (new_width - new_img_width) // 2 + paste_y = (new_height - new_img_height) // 2 + + new_image.paste(resized_img, (paste_x, paste_y)) + + return new_image + +def resize_image_with_target_area(img: Image.Image, target_area: int = 1024 * 1024) -> Image.Image: + """ + 将 PIL 图像缩放到接近指定像素面积（target_area），保持原始宽高比， + 并确保新宽度和高度均为 32 的整数倍。 + + 参数: + img (PIL.Image.Image): 输入图像 + target_area (int): 目标像素总面积，例如 1024*1024 = 1048576 + + 返回: + PIL.Image.Image: Resize 后的图像 + """ + orig_w, orig_h = img.size + if orig_w == 0 or orig_h == 0: + raise ValueError("Input image has zero width or height.") + + ratio = orig_w / orig_h + ideal_width = math.sqrt(target_area * ratio) + ideal_height = ideal_width / ratio + + new_width = round(ideal_width / 32) * 32 + new_height = round(ideal_height / 32) * 32 + + new_width = max(32, new_width) + new_height = max(32, new_height) + + new_width = int(new_width) + new_height = int(new_height) + + resized_img = img.resize((new_width, new_height), Image.LANCZOS) + return resized_img + +class Camera(object): + """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + def __init__(self, entry): + fx, fy, cx, cy = entry[1:5] + self.fx = fx + self.fy = fy + self.cx = cx + self.cy = cy + w2c_mat = np.array(entry[7:]).reshape(3, 4) + w2c_mat_4x4 = np.eye(4) + w2c_mat_4x4[:3, :] = w2c_mat + self.w2c_mat = w2c_mat_4x4 + self.c2w_mat = np.linalg.inv(w2c_mat_4x4) + +def custom_meshgrid(*args): + """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + # ref: https://pytorch.org/docs/stable/generated/torch.meshgrid.html?highlight=meshgrid#torch.meshgrid + if pver.parse(torch.__version__) < pver.parse('1.10'): + return torch.meshgrid(*args) + else: + return torch.meshgrid(*args, indexing='ij') + +def get_relative_pose(cam_params): + """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + abs_w2cs = [cam_param.w2c_mat for cam_param in cam_params] + abs_c2ws = [cam_param.c2w_mat for cam_param in cam_params] + cam_to_origin = 0 + target_cam_c2w = np.array([ + [1, 0, 0, 0], + [0, 1, 0, -cam_to_origin], + [0, 0, 1, 0], + [0, 0, 0, 1] + ]) + abs2rel = target_cam_c2w @ abs_w2cs[0] + ret_poses = [target_cam_c2w, ] + [abs2rel @ abs_c2w for abs_c2w in abs_c2ws[1:]] + ret_poses = np.array(ret_poses, dtype=np.float32) + return ret_poses + +def ray_condition(K, c2w, H, W, device): + """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + # c2w: B, V, 4, 4 + # K: B, V, 4 + + B = K.shape[0] + + j, i = custom_meshgrid( + torch.linspace(0, H - 1, H, device=device, dtype=c2w.dtype), + torch.linspace(0, W - 1, W, device=device, dtype=c2w.dtype), + ) + i = i.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5 # [B, HxW] + j = j.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5 # [B, HxW] + + fx, fy, cx, cy = K.chunk(4, dim=-1) # B,V, 1 + + zs = torch.ones_like(i) # [B, HxW] + xs = (i - cx) / fx * zs + ys = (j - cy) / fy * zs + zs = zs.expand_as(ys) + + directions = torch.stack((xs, ys, zs), dim=-1) # B, V, HW, 3 + directions = directions / directions.norm(dim=-1, keepdim=True) # B, V, HW, 3 + + rays_d = directions @ c2w[..., :3, :3].transpose(-1, -2) # B, V, 3, HW + rays_o = c2w[..., :3, 3] # B, V, 3 + rays_o = rays_o[:, :, None].expand_as(rays_d) # B, V, 3, HW + # c2w @ dirctions + rays_dxo = torch.cross(rays_o, rays_d) + plucker = torch.cat([rays_dxo, rays_d], dim=-1) + plucker = plucker.reshape(B, c2w.shape[1], H, W, 6) # B, V, H, W, 6 + # plucker = plucker.permute(0, 1, 4, 2, 3) + return plucker + +def process_pose_file(pose_file_path, width=672, height=384, original_pose_width=1280, original_pose_height=720, device='cpu', return_poses=False): + """Modified from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + with open(pose_file_path, 'r') as f: + poses = f.readlines() + + poses = [pose.strip().split(' ') for pose in poses[1:]] + cam_params = [[float(x) for x in pose] for pose in poses] + if return_poses: + return cam_params + else: + cam_params = [Camera(cam_param) for cam_param in cam_params] + + sample_wh_ratio = width / height + pose_wh_ratio = original_pose_width / original_pose_height # Assuming placeholder ratios, change as needed + + if pose_wh_ratio > sample_wh_ratio: + resized_ori_w = height * pose_wh_ratio + for cam_param in cam_params: + cam_param.fx = resized_ori_w * cam_param.fx / width + else: + resized_ori_h = width / pose_wh_ratio + for cam_param in cam_params: + cam_param.fy = resized_ori_h * cam_param.fy / height + + intrinsic = np.asarray([[cam_param.fx * width, + cam_param.fy * height, + cam_param.cx * width, + cam_param.cy * height] + for cam_param in cam_params], dtype=np.float32) + + K = torch.as_tensor(intrinsic)[None] # [1, 1, 4] + c2ws = get_relative_pose(cam_params) # Assuming this function is defined elsewhere + c2ws = torch.as_tensor(c2ws)[None] # [1, n_frame, 4, 4] + plucker_embedding = ray_condition(K, c2ws, height, width, device=device)[0].permute(0, 3, 1, 2).contiguous() # V, 6, H, W + plucker_embedding = plucker_embedding[None] + plucker_embedding = rearrange(plucker_embedding, "b f c h w -> b f h w c")[0] + return plucker_embedding + +def process_pose_params(cam_params, width=672, height=384, original_pose_width=1280, original_pose_height=720, device='cpu'): + """Modified from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + cam_params = [Camera(cam_param) for cam_param in cam_params] + + sample_wh_ratio = width / height + pose_wh_ratio = original_pose_width / original_pose_height # Assuming placeholder ratios, change as needed + + if pose_wh_ratio > sample_wh_ratio: + resized_ori_w = height * pose_wh_ratio + for cam_param in cam_params: + cam_param.fx = resized_ori_w * cam_param.fx / width + else: + resized_ori_h = width / pose_wh_ratio + for cam_param in cam_params: + cam_param.fy = resized_ori_h * cam_param.fy / height + + intrinsic = np.asarray([[cam_param.fx * width, + cam_param.fy * height, + cam_param.cx * width, + cam_param.cy * height] + for cam_param in cam_params], dtype=np.float32) + + K = torch.as_tensor(intrinsic)[None] # [1, 1, 4] + c2ws = get_relative_pose(cam_params) # Assuming this function is defined elsewhere + c2ws = torch.as_tensor(c2ws)[None] # [1, n_frame, 4, 4] + plucker_embedding = ray_condition(K, c2ws, height, width, device=device)[0].permute(0, 3, 1, 2).contiguous() # V, 6, H, W + plucker_embedding = plucker_embedding[None] + plucker_embedding = rearrange(plucker_embedding, "b f c h w -> b f h w c")[0] + return plucker_embedding \ No newline at end of file diff --git a/videox_fun/models/fantasytalking_audio_encoder.py b/videox_fun/models/fantasytalking_audio_encoder.py index c3be73e..7f2f760 100644 --- a/videox_fun/models/fantasytalking_audio_encoder.py +++ b/videox_fun/models/fantasytalking_audio_encoder.py @@ -38,6 +38,15 @@ class FantasyTalkingAudioEncoder(ModelMixin, ConfigMixin, FromOriginalModelMixin audio_segment, sampling_rate=sample_rate, return_tensors="pt" ).input_values.to(self.model.device, self.model.dtype) + with torch.no_grad(): + fea = self.model(input_values).last_hidden_state + return fea + + def extract_audio_feat_without_file_load(self, audio_segment, sample_rate): + input_values = self.processor( + audio_segment, sampling_rate=sample_rate, return_tensors="pt" + ).input_values.to(self.model.device, self.model.dtype) + with torch.no_grad(): fea = self.model(input_values).last_hidden_state return fea \ No newline at end of file diff --git a/videox_fun/pipeline/pipeline_fantasy_talking.py b/videox_fun/pipeline/pipeline_fantasy_talking.py index f0f03ba..114bd22 100644 --- a/videox_fun/pipeline/pipeline_fantasy_talking.py +++ b/videox_fun/pipeline/pipeline_fantasy_talking.py @@ -695,7 +695,7 @@ class FantasyTalkingPipeline(DiffusionPipeline): ) audio_scale = torch.tensor( - [0, 1] + [0.75, 1] ).to(latent_model_input.device, latent_model_input.dtype) # broadcast to batch dimension in a way that's compatible with ONNX/Core ML