# coding=utf-8 # Copyright 2025 The Kwai Keye Team and The HuggingFace Inc. team. All rights reserved. # # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX # and OPT implementations in this library. It has been modified from its # original forms to accommodate minor architectural differences compared # to GPT-NeoX and OPT used by the Meta AI team that trained the model. # # 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 List, Union, Optional from transformers.feature_extraction_utils import BatchFeature from transformers.image_utils import ImageInput from transformers.video_utils import VideoInput from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack, VideosKwargs from transformers.tokenization_utils_base import PreTokenizedInput, TextInput from .image_processing_keye_vl_1_5 import KeyeVL1_5ImageProcessor import torch import torch.nn as nn import numpy as np from itertools import chain class KeyeVL1_5VideosProcessorKwargs(VideosKwargs, total=False): fps: Optional[Union[List[float], float]] # 准备reszie到的width（slow） width: Optional[Union[List[int], int]] # 准备reszie到的height（slow） height: Optional[Union[List[int], int]] # 准备resize到的width（fast） fast_width: Optional[Union[List[int], int]] # 准备resize到的height（fast） fast_height: Optional[Union[List[int], int]] # 用于标记每一帧的时间戳，数量和帧数相等 timestamps: Optional[Union[List[torch.Tensor], torch.Tensor]] # 用于标记每一帧的类型是slow还是fast，slow=0, fast=1 frame_types: Optional[Union[List[torch.Tensor], torch.Tensor]] class KeyeVL1_5ProcessorKwargs(ProcessingKwargs, total=False): videos_kwargs: KeyeVL1_5VideosProcessorKwargs _defaults = { "text_kwargs": { "padding": False, }, "videos_kwargs": {"fps": 2.0}, } def select_slow_fast_frames(frames: torch.Tensor, frame_types: torch.Tensor): """ Selects frames from a tensor based on a mask list. Args: frames (torch.Tensor): A tensor of shape (nframes, c, h, w). frame_types (torch.Tensor): A int tensor of shape (nframes,) Returns: tuple[torch.Tensor, torch.Tensor]: A tuple containing two tensors: - slow_frames: Frames which the type is 0. - fast_frames: Frames where the type is 1. """ nframes, _, _, _ = frames.shape if frame_types.shape[-1] != nframes: raise ValueError("Length of mask must be equal to the number of frames.") mask = (frame_types == 0) slow_frames = frames[mask] fast_frames = frames[~mask] return slow_frames, fast_frames def split_thw(tensor): """Split grid_thw in t dimension, the result tensor should like [[1, h, w],...]""" repeats = tensor[:, 0] new_thw = torch.cat([ torch.ones(tensor.shape[0], 1, dtype=tensor.dtype, device=tensor.device), tensor[:, 1:] ], dim=1) return torch.repeat_interleave(new_thw, repeats, dim=0) def merge_hws(hws): """ 优化版本：使用更高效的方法合并张量 """ merged = [] last_hw = [-1, -1] for hw in hws: # 找到连续相同形状的张量 if hw[1:] == last_hw: merged[-1][0] += 1 else: merged.append(hw) last_hw = hw[1:] return torch.tensor(merged) class KeyeVL1_5Processor(ProcessorMixin): r""" [`KeyeVL1_5Processor`] offers all the functionalities of [`KeyeVL1_5ImageProcessor`] and [`Qwen2TokenizerFast`]. See the [`~KeyeVL1_5Processor.__call__`] and [`~KeyeVL1_5Processor.decode`] for more information. Args: image_processor ([`KeyeVL1_5ImageProcessor`], *optional*): The image processor is a required input. tokenizer ([`Qwen2TokenizerFast`], *optional*): The tokenizer is a required input. chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string. """ attributes = ["image_processor", "tokenizer"] valid_kwargs = [ "chat_template","image_std", "min_pixels", "image_mean", "merge_size", "image_processor_type", "temporal_patch_size", "patch_size", "max_pixels" ] image_processor_class = "AutoImageProcessor" tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast") def __init__(self, image_processor=None, tokenizer=None, chat_template=None, **kwargs): self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token self.video_token = "<|video_pad|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token self.frame_token = "<|frame|>" if not hasattr(tokenizer, "frame_token") else tokenizer.frame_token self.fast_video_token = "<|fast_video_pad|>" if not hasattr(tokenizer, "fast_video_token") else tokenizer.fast_video_token self.fast_start = "<|fast_start|>" if not hasattr(tokenizer, "fast_start") else tokenizer.fast_start self.fast_end = "<|fast_end|>" if not hasattr(tokenizer, "fast_end") else tokenizer.fast_end super().__init__(image_processor, tokenizer, chat_template=chat_template) self.slowfast = True def __call__( self, text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None, images: ImageInput = None, videos: VideoInput = None, **kwargs: Unpack[KeyeVL1_5ProcessorKwargs], ) -> BatchFeature: """ Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text` and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to KeyeVL1_5ImageProcessor's [`~KeyeVL1_5ImageProcessor.__call__`] if `vision_infos` is not `None`. Args: text (`str`, `List[str]`, `List[List[str]]`): The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set `is_split_into_words=True` (to lift the ambiguity with a batch of sequences). images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`): The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch tensor. Both channels-first and channels-last formats are supported. videos (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`): The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported. return_tensors (`str` or [`~utils.TensorType`], *optional*): If set, will return tensors of a particular framework. Acceptable values are: - `'tf'`: Return TensorFlow `tf.constant` objects. - `'pt'`: Return PyTorch `torch.Tensor` objects. - `'np'`: Return NumPy `np.ndarray` objects. - `'jax'`: Return JAX `jnp.ndarray` objects. Returns: [`BatchFeature`]: A [`BatchFeature`] with the following fields: - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`. - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not `None`). - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`. - **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`. - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`. - **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`. - **second_per_grid_ts** -- List of video seconds per time grid. Returned when `videos` is not `None`. """ output_kwargs = self._merge_kwargs( KeyeVL1_5ProcessorKwargs, tokenizer_init_kwargs=self.tokenizer.init_kwargs, **kwargs, ) if images is not None: # slow_images = images image_inputs = self.image_processor(images=images, return_tensors="pt") image_inputs['pixel_values'] = image_inputs['pixel_values'] image_grid_thw = image_inputs["image_grid_thw"] else: image_inputs = {} image_grid_thw = None num_frames = [] if videos is not None: batch_slow_frames = [] batch_fast_frames = [] videos_kwargs = output_kwargs["videos_kwargs"] num_videos = len(videos) batch_frame_types = videos_kwargs.get("frame_types", [None] * num_videos) batch_timestamps = videos_kwargs.get("timestamps", [None] * num_videos) batch_width = videos_kwargs.get("width", [None] * num_videos) batch_height = videos_kwargs.get("height", [None] * num_videos) batch_fast_width = videos_kwargs.get("fast_width", [None] * num_videos) batch_fast_height = videos_kwargs.get("fast_height", [None] * num_videos) for index, frames in enumerate(videos): if isinstance(frames, np.ndarray): frames = torch.from_numpy(frames) nframes = frames.shape[0] num_frames.append(nframes) assert nframes > 0, "No frames in video" if batch_frame_types[index] is None: # default to all slow frames batch_frame_types[index] = torch.zeros((nframes, ), dtype=torch.long) frame_types = batch_frame_types[index] slow_frames, fast_frames = select_slow_fast_frames(frames, frame_types) has_fast_frames = fast_frames.shape[0] > 0 # resize slow frames resized_width = batch_width[index] resized_height = batch_height[index] if resized_width is not None and resized_height is not None: slow_frames = nn.functional.interpolate( slow_frames, [resized_height, resized_width], mode="bilinear", antialias=True, ).float() do_resize = False else: slow_frames = slow_frames.float() do_resize = True # Tensor(N, C, H, W) -> Tuple[Tensor(1, C, H, W)] # slow_frames = list(slow_frames.split(1, dim=0))，不split，在模型里面做 slow_video_inputs = self.image_processor( images=None, videos=[slow_frames], **output_kwargs["images_kwargs"], do_resize=do_resize) slow_video_grid_thw = slow_video_inputs["video_grid_thw"] batch_slow_frames.append(slow_video_inputs) # # 当前这个视频每一帧的token数 # slow_frames_patch_nums[index] = int(slow_video_inputs["pixel_values_videos"].shape[0] / \ # slow_video_grid_thw.squeeze()[0]) if has_fast_frames: # TODO: shrink fast_frames fast_resized_width = batch_fast_width[index] fast_resized_height = batch_fast_height[index] if fast_resized_width is not None and fast_resized_height is not None: fast_frames = nn.functional.interpolate( fast_frames, [fast_resized_height, fast_resized_width], mode="bilinear", antialias=True, ).float() do_fast_resize = False else: fast_frames = fast_frames.float() do_fast_resize = True # Tensor(N, C, H, W) -> Tuple[Tensor(1, C, H, W)] # fast_frames = list(fast_frames.split(1, dim=0)) fast_video_inputs = self.image_processor( images=None, videos=[fast_frames], **output_kwargs["images_kwargs"], do_resize=do_fast_resize) fast_video_grid_thw = fast_video_inputs["video_grid_thw"] batch_fast_frames.append(fast_video_inputs) # # 当前这个视频的所有token数 # fast_frames_token_nums[index] = int(fast_video_inputs["pixel_values_videos"].shape[0] / \ # fast_video_grid_thw.squeeze()[0]) assert len(batch_slow_frames) > 0, "Slow frames should not be empty." slow_pixel_values_videos_list = [ video["pixel_values_videos"] for video in batch_slow_frames if video is not None] slow_video_grid_thw_list = [ video["video_grid_thw"] for video in batch_slow_frames if video is not None] slow_pixel_values_videos = torch.concat(slow_pixel_values_videos_list, dim=0) slow_video_grid_thw = torch.concat(slow_video_grid_thw_list, dim=0) if has_fast_frames: fast_pixel_values_videos_list = [ video["pixel_values_videos"] for video in batch_fast_frames \ if video is not None] fast_video_grid_thw_list = [ video["video_grid_thw"] for video in batch_fast_frames \ if video is not None] fast_pixel_values_videos = \ torch.concat(fast_pixel_values_videos_list, dim=0) fast_video_grid_thw = \ torch.concat(fast_video_grid_thw_list, dim=0) else: fast_video_grid_thw = None else: slow_video_grid_thw = None fast_video_grid_thw = None if not isinstance(text, list): text = [text] if image_grid_thw is not None: index = 0 for i in range(len(text)): while self.image_token in text[i]: image_place_holder_tempale = "<|placeholder|>" * ( image_grid_thw[index].prod() // self.image_processor.merge_size ** 2) text[i] = text[i].replace( self.image_token, image_place_holder_tempale, 1, ) index += 1 text[i] = text[i].replace("<|placeholder|>", self.image_token) pixel_values_videos = [] video_grid_thw = [] videos_inputs = {} if slow_video_grid_thw is not None: slow_video_grid_thw = split_thw(slow_video_grid_thw) if fast_video_grid_thw is not None: fast_video_grid_thw = split_thw(fast_video_grid_thw) index = 0 slow_index = 0 fast_index = 0 slow_pixels_index = 0 fast_pixels_index = 0 for i in range(len(text)): while self.video_token in text[i]: video_place_holder_tempale = "" for j in range(batch_frame_types[index].shape[-1]): if batch_timestamps[index] is not None: # 如果有时间戳 video_place_holder_tempale += self.frame_token + format(batch_timestamps[index][j], ".1f") else: video_place_holder_tempale += self.frame_token # 当前帧是slow if batch_frame_types[index][j] == 0: num_patches = int(slow_video_grid_thw[slow_index].prod()) video_place_holder_tempale += "<|placeholder|>" * ( num_patches // self.image_processor.merge_size ** 2) pixel_values_videos.append( slow_pixel_values_videos[slow_pixels_index:slow_pixels_index + num_patches]) slow_pixels_index = slow_pixels_index + num_patches video_grid_thw.append(slow_video_grid_thw[slow_index].tolist()) slow_index += 1 # 当前帧是fast elif batch_frame_types[index][j] == 1: num_patches = int(fast_video_grid_thw[fast_index].prod()) video_place_holder_tempale += self.fast_start + "<|placeholder|>" * ( num_patches // self.image_processor.merge_size ** 2) + \ self.fast_end pixel_values_videos.append( fast_pixel_values_videos[fast_pixels_index:fast_pixels_index + num_patches]) fast_pixels_index = fast_pixels_index + num_patches video_grid_thw.append(fast_video_grid_thw[fast_index].tolist()) fast_index += 1 text[i] = text[i].replace( self.video_token, video_place_holder_tempale, 1, ) index += 1 text[i] = text[i].replace("<|placeholder|>", self.video_token) videos_inputs["pixel_values_videos"] = torch.cat(pixel_values_videos, dim=0) videos_inputs["video_grid_thw"] = merge_hws(video_grid_thw) videos_inputs["num_frames"] = torch.tensor(num_frames) text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"]) return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs}) def batch_decode(self, *args, **kwargs): """ This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please refer to the docstring of this method for more information. """ return self.tokenizer.batch_decode(*args, **kwargs) def decode(self, *args, **kwargs): """ This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to the docstring of this method for more information. """ return self.tokenizer.decode(*args, **kwargs) def post_process_image_text_to_text( self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs ): """ Post-process the output of the model to decode the text. Args: generated_outputs (`torch.Tensor` or `np.ndarray`): The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)` or `(sequence_length,)`. skip_special_tokens (`bool`, *optional*, defaults to `True`): Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method. Clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`): Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method. **kwargs: Additional arguments to be passed to the tokenizer's `batch_decode method`. Returns: `List[str]`: The decoded text. """ return self.tokenizer.batch_decode( generated_outputs, skip_special_tokens=skip_special_tokens, clean_up_tokenization_spaces=clean_up_tokenization_spaces, **kwargs, ) @property def model_input_names(self): tokenizer_input_names = self.tokenizer.model_input_names image_processor_input_names = self.image_processor.model_input_names names_from_processor = list(dict.fromkeys(tokenizer_input_names + image_processor_input_names)) return names_from_processor __all__ = ["KeyeVL1_5Processor"]