205 lines
9.3 KiB
Python
205 lines
9.3 KiB
Python
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates and/or its affiliates
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
|
|
from typing import TYPE_CHECKING, ClassVar#, Unpack
|
|
|
|
import numpy as np
|
|
from loguru import logger
|
|
from transformers.image_utils import ImageInput
|
|
from transformers.models.qwen2_5_vl.processing_qwen2_5_vl import (
|
|
Qwen2_5_VLImagesKwargs,
|
|
Qwen2_5_VLProcessor,
|
|
Qwen2_5_VLVideosProcessorKwargs,
|
|
)
|
|
from transformers.processing_utils import BatchFeature, PreTokenizedInput, ProcessingKwargs, TextInput
|
|
from transformers.video_utils import VideoInput
|
|
|
|
from .mammothmoda2_qwen2_5_vl import MammothUTokenizer
|
|
|
|
if TYPE_CHECKING:
|
|
from transformers.models.qwen2_vl import Qwen2VLImageProcessor, Qwen2VLImageProcessorFast
|
|
from transformers.models.qwen2_vl.video_processing_qwen2_vl import Qwen2VLVideoProcessor
|
|
|
|
|
|
DEFAULT_NEGATIVE_PROMPT = (
|
|
"deformed, blurry, over saturation, bad anatomy, disfigured, poorly drawn face, mutation, mutated, "
|
|
"extra_limb, ugly, poorly drawn hands, fused fingers, messy drawing, broken legs censor, censored, "
|
|
"censor_bar Create an image from the instruction."
|
|
)
|
|
|
|
|
|
class Mammothmoda2ImagesKwargs(Qwen2_5_VLImagesKwargs):
|
|
negative_prompt: str | None
|
|
num_images_per_prompt: int
|
|
cfg_scale: float
|
|
|
|
|
|
class Mammothmoda2ProcessorKwargs(ProcessingKwargs, total=False):
|
|
images_kwargs: Mammothmoda2ImagesKwargs
|
|
videos_kwargs: Qwen2_5_VLVideosProcessorKwargs
|
|
_defaults = { # noqa: RUF012
|
|
"text_kwargs": {
|
|
"padding": False,
|
|
"return_mm_token_type_ids": False,
|
|
},
|
|
}
|
|
|
|
|
|
class Mammothmoda2Processor(Qwen2_5_VLProcessor):
|
|
"""The mammothmoda2 processor inherit from Qwen2_5_VLProcessor, adding image editing support."""
|
|
|
|
attributes: ClassVar[list[str]] = ["image_processor", "tokenizer", "video_processor"]
|
|
|
|
image_processor_class = "AutoImageProcessor"
|
|
video_processor_class = "AutoVideoProcessor"
|
|
tokenizer_class = "MammothUTokenizer"
|
|
|
|
def __init__(
|
|
self,
|
|
image_processor=None,
|
|
tokenizer: MammothUTokenizer | None = None,
|
|
video_processor=None,
|
|
chat_template=None,
|
|
t2i_generate: bool = False,
|
|
ar_height: int = 32,
|
|
ar_width: int = 32,
|
|
**kwargs, # noqa: ARG002
|
|
) -> None:
|
|
super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)
|
|
self.t2i_generate = t2i_generate
|
|
self.ar_height = ar_height
|
|
self.ar_width = ar_width
|
|
logger.info(f"Mammothmoda2Processor init: {t2i_generate=} | {ar_height=} | {ar_width=}")
|
|
|
|
# Type maker for better IDE type hint.
|
|
self.tokenizer: MammothUTokenizer
|
|
self.image_processor: Qwen2VLImageProcessor | Qwen2VLImageProcessorFast
|
|
self.video_processor: Qwen2VLVideoProcessor
|
|
|
|
def __call__(
|
|
self,
|
|
images: ImageInput | None = None,
|
|
text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None = None,
|
|
videos: VideoInput | None = None,
|
|
# **kwargs: Unpack[Mammothmoda2ProcessorKwargs],
|
|
**kwargs,
|
|
) -> BatchFeature:
|
|
output_kwargs = self._merge_kwargs(
|
|
Mammothmoda2ProcessorKwargs,
|
|
tokenizer_init_kwargs=self.tokenizer.init_kwargs,
|
|
**kwargs,
|
|
)
|
|
if not isinstance(text, list):
|
|
text = [text]
|
|
|
|
# Mammothmoda2 pre-processing: inputs expansion.
|
|
if self.t2i_generate is True:
|
|
num_images_per_prompt = output_kwargs["images_kwargs"]["num_images_per_prompt"]
|
|
cfg_scale = output_kwargs["images_kwargs"]["cfg_scale"]
|
|
if num_images_per_prompt > 1: # NOTE: num_images_per_prompt > 1, we need to repeat the inputs
|
|
images = images * num_images_per_prompt if images is not None else None
|
|
videos = videos * num_images_per_prompt if videos is not None else None
|
|
text = text * num_images_per_prompt
|
|
if cfg_scale > 1.0: # NOTE: cfg_scale > 1.0, we need to repeat the inputs
|
|
images = images * 2 if images is not None else None
|
|
videos = videos * 2 if videos is not None else None
|
|
text = text * 2
|
|
|
|
# Original Qwen2_5_VLProcessor logic.
|
|
image_inputs = videos_inputs = {}
|
|
if images is not None:
|
|
image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"])
|
|
image_grid_thw = image_inputs["image_grid_thw"]
|
|
if videos is not None:
|
|
fps = output_kwargs["videos_kwargs"].get("fps", 2.0)
|
|
videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"])
|
|
video_grid_thw = videos_inputs["video_grid_thw"]
|
|
|
|
if isinstance(fps, (int, float)):
|
|
second_per_grid_ts = [self.video_processor.temporal_patch_size / fps] * len(video_grid_thw)
|
|
elif hasattr(fps, "__len__") and len(fps) == len(video_grid_thw):
|
|
second_per_grid_ts = [self.video_processor.temporal_patch_size / tmp for tmp in fps]
|
|
else:
|
|
error_msg = (
|
|
f"The length of fps ({len(fps) if hasattr(fps, '__len__') else fps}) must be equal to the "
|
|
f"length of video_grid_thw ({len(video_grid_thw)}) or fps should be a single number."
|
|
)
|
|
raise ValueError(error_msg)
|
|
videos_inputs.update({"second_per_grid_ts": second_per_grid_ts})
|
|
|
|
text = text.copy() # below lines change text in-place
|
|
if images is not None:
|
|
merge_length = self.image_processor.merge_size**2
|
|
index = 0
|
|
for i in range(len(text)):
|
|
while self.image_token in text[i]:
|
|
num_image_tokens = image_grid_thw[index].prod() // merge_length
|
|
text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
|
|
index += 1
|
|
text[i] = text[i].replace("<|placeholder|>", self.image_token)
|
|
|
|
if videos is not None:
|
|
merge_length = self.video_processor.merge_size**2
|
|
index = 0
|
|
for i in range(len(text)):
|
|
while self.video_token in text[i]:
|
|
num_video_tokens = video_grid_thw[index].prod() // merge_length
|
|
text[i] = text[i].replace(self.video_token, "<|placeholder|>" * num_video_tokens, 1)
|
|
index += 1
|
|
text[i] = text[i].replace("<|placeholder|>", self.video_token)
|
|
|
|
return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
|
|
return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", None)
|
|
text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
|
|
self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"])
|
|
|
|
if return_mm_token_type_ids:
|
|
array_ids = np.array(text_inputs["input_ids"])
|
|
mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
|
|
mm_token_type_ids[array_ids == self.image_token_id] = 1
|
|
text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
|
|
|
|
inputs = BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs}, tensor_type=return_tensors)
|
|
|
|
# Mammothmoda2 t2i post-processing: attaching negative prompt.
|
|
if self.t2i_generate is True:
|
|
negative_ids, negative_mask = None, None
|
|
if (negative_prompt := output_kwargs["images_kwargs"].get("negative_prompt", None)) is not None:
|
|
negative_messages = [
|
|
{"role": "system", "content": [{"type": "text", "text": "You are a helpful image generator."}]},
|
|
{"role": "user", "content": [{"type": "text", "text": negative_prompt}]},
|
|
]
|
|
negative_text = self.apply_chat_template(negative_messages, tokenize=False, add_generation_prompt=False)
|
|
negative_inputs = super().__call__(
|
|
text=[negative_text] * num_images_per_prompt,
|
|
images=None,
|
|
videos=None,
|
|
return_tensors=return_tensors,
|
|
padding=True,
|
|
padding_side="left",
|
|
)
|
|
negative_ids = negative_inputs.input_ids # [bs, seq_len]
|
|
negative_mask = negative_inputs.attention_mask # full 1
|
|
inputs["negative_ids"] = negative_ids # Already Tensor, directly attach.
|
|
inputs["negative_mask"] = negative_mask
|
|
return inputs
|
|
|
|
def apply_chat_template(self, *args, **kwargs) -> str:
|
|
if self.t2i_generate is True: # For t2i, use different chat template.
|
|
kwargs["t2i_generate"] = self.t2i_generate
|
|
kwargs["ar_height"] = self.ar_height
|
|
kwargs["ar_width"] = self.ar_width
|
|
return super().apply_chat_template(*args, **kwargs)
|