Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.3#4# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX5# and OPT implementations in this library. It has been modified from its6# original forms to accommodate minor architectural differences compared7# to GPT-NeoX and OPT used by the Meta AI team that trained the model.8#9# Licensed under the Apache License, Version 2.0 (the "License");10# you may not use this file except in compliance with the License.11# You may obtain a copy of the License at12#13#     http://www.apache.org/licenses/LICENSE-2.014#15# Unless required by applicable law or agreed to in writing, software16# distributed under the License is distributed on an "AS IS" BASIS,17# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.18# See the License for the specific language governing permissions and19# limitations under the License.20"""21Processor class for Qwen2-VL.22"""23 24from typing import Optional, Union25 26import numpy as np27 28from ...feature_extraction_utils import BatchFeature29from ...image_utils import ImageInput30from ...processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack31from ...tokenization_utils_base import PreTokenizedInput, TextInput32from ...utils import logging33from ...video_utils import VideoInput34 35 36logger = logging.get_logger(__name__)37 38 39class Qwen2VLImagesKwargs(ImagesKwargs):40    min_pixels: Optional[int]41    max_pixels: Optional[int]42    patch_size: Optional[int]43    temporal_patch_size: Optional[int]44    merge_size: Optional[int]45 46 47class Qwen2VLProcessorKwargs(ProcessingKwargs, total=False):48    images_kwargs: Qwen2VLImagesKwargs49    _defaults = {50        "text_kwargs": {51            "padding": False,52            "return_mm_token_type_ids": False,53        },54    }55 56 57class Qwen2VLProcessor(ProcessorMixin):58    r"""59    Constructs a Qwen2-VL processor which wraps a Qwen2-VL image processor and a Qwen2 tokenizer into a single processor.60    [`Qwen2VLProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the61    [`~Qwen2VLProcessor.__call__`] and [`~Qwen2VLProcessor.decode`] for more information.62    Args:63        image_processor ([`Qwen2VLImageProcessor`], *optional*):64            The image processor is a required input.65        tokenizer ([`Qwen2TokenizerFast`], *optional*):66            The tokenizer is a required input.67        video_processor ([`Qwen2VLVideoProcessor`], *optional*):68            The video processor is a required input.69        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages70            in a chat into a tokenizable string.71    """72 73    attributes = ["image_processor", "tokenizer", "video_processor"]74    image_processor_class = "AutoImageProcessor"75    video_processor_class = "AutoVideoProcessor"76    tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")77 78    def __init__(self, image_processor=None, tokenizer=None, video_processor=None, chat_template=None, **kwargs):79        self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token80        self.video_token = "<|video_pad|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token81        self.image_token_id = (82            tokenizer.image_token_id83            if getattr(tokenizer, "image_token_id", None)84            else tokenizer.convert_tokens_to_ids(self.image_token)85        )86        self.video_token_id = (87            tokenizer.video_token_id88            if getattr(tokenizer, "video_token_id", None)89            else tokenizer.convert_tokens_to_ids(self.video_token)90        )91        super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)92 93    def __call__(94        self,95        images: Optional[ImageInput] = None,96        text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,97        videos: Optional[VideoInput] = None,98        **kwargs: Unpack[Qwen2VLProcessorKwargs],99    ) -> BatchFeature:100        """101        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`102        and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode103        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwargs` arguments to104        Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.105 106        Args:107            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):108                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch109                tensor. Both channels-first and channels-last formats are supported.110            text (`str`, `list[str]`, `list[list[str]]`):111                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings112                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set113                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).114            videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):115                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch116                tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.117            return_tensors (`str` or [`~utils.TensorType`], *optional*):118                If set, will return tensors of a particular framework. Acceptable values are:119                - `'tf'`: Return TensorFlow `tf.constant` objects.120                - `'pt'`: Return PyTorch `torch.Tensor` objects.121                - `'np'`: Return NumPy `np.ndarray` objects.122                - `'jax'`: Return JAX `jnp.ndarray` objects.123 124        Returns:125            [`BatchFeature`]: A [`BatchFeature`] with the following fields:126 127            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.128            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when129              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not130              `None`).131            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.132            - **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`.133            - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`.134            - **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`.135        """136        output_kwargs = self._merge_kwargs(137            Qwen2VLProcessorKwargs,138            tokenizer_init_kwargs=self.tokenizer.init_kwargs,139            **kwargs,140        )141 142        image_inputs = videos_inputs = {}143        if images is not None:144            image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"])145            image_grid_thw = image_inputs["image_grid_thw"]146 147        if videos is not None:148            videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"])149            video_grid_thw = videos_inputs["video_grid_thw"]150 151        if not isinstance(text, list):152            text = [text]153 154        text = text.copy()  # below lines change text in-place155 156        if images is not None:157            merge_length = self.image_processor.merge_size**2158            index = 0159            for i in range(len(text)):160                while self.image_token in text[i]:161                    num_image_tokens = image_grid_thw[index].prod() // merge_length162                    text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)163                    index += 1164                text[i] = text[i].replace("<|placeholder|>", self.image_token)165 166        if videos is not None:167            merge_length = self.video_processor.merge_size**2168            index = 0169            for i in range(len(text)):170                while self.video_token in text[i]:171                    num_video_tokens = video_grid_thw[index].prod() // merge_length172                    text[i] = text[i].replace(self.video_token, "<|placeholder|>" * num_video_tokens, 1)173                    index += 1174                text[i] = text[i].replace("<|placeholder|>", self.video_token)175 176        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)177        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", False)178        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"], return_tensors=None)179        self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"])180 181        if return_mm_token_type_ids:182            array_ids = np.array(text_inputs["input_ids"])183            mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])184            mm_token_type_ids[array_ids == self.image_token_id] = 1185            text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()186 187        return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs}, tensor_type=return_tensors)188 189    def _get_num_multimodal_tokens(self, image_sizes=None, video_sizes=None, **kwargs):190        """191        Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.192        Args:193            image_sizes (`list[list[int]]`, *optional*):194                The input sizes formatted as (height, width) per each image.195            video_sizes (`list[list[int]]`, *optional*):196                The input sizes formatted as (num_frames, height, width) per each video.197        Returns:198            `MultiModalData`: A `MultiModalData` object holding number of tokens per each of the provided199            input modalities, along with other useful data.200        """201 202        vision_data = {}203        if image_sizes is not None:204            images_kwargs = Qwen2VLProcessorKwargs._defaults.get("images_kwargs", {})205            images_kwargs.update(kwargs)206            merge_size = images_kwargs.get("merge_size", None) or self.image_processor.merge_size207 208            num_image_patches = [209                self.image_processor.get_number_of_image_patches(*image_size, images_kwargs)210                for image_size in image_sizes211            ]212            num_image_tokens = [(num_patches // merge_size**2) for num_patches in num_image_patches]213            vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches})214 215        if video_sizes is not None:216            videos_kwargs = Qwen2VLProcessorKwargs._defaults.get("videos_kwargs", {})217            videos_kwargs.update(kwargs)218            num_video_patches = [219                self.video_processor.get_number_of_video_patches(*video_size, videos_kwargs)220                for video_size in video_sizes221            ]222            num_video_tokens = [(num_patches // merge_size**2) for num_patches in num_video_patches]223            vision_data["num_video_tokens"] = num_video_tokens224 225        return MultiModalData(**vision_data)226 227    def post_process_image_text_to_text(228        self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs229    ):230        """231        Post-process the output of the model to decode the text.232 233        Args:234            generated_outputs (`torch.Tensor` or `np.ndarray`):235                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`236                or `(sequence_length,)`.237            skip_special_tokens (`bool`, *optional*, defaults to `True`):238                Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method.239            clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):240                Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method.241            **kwargs:242                Additional arguments to be passed to the tokenizer's `batch_decode method`.243 244        Returns:245            `list[str]`: The decoded text.246        """247        return self.tokenizer.batch_decode(248            generated_outputs,249            skip_special_tokens=skip_special_tokens,250            clean_up_tokenization_spaces=clean_up_tokenization_spaces,251            **kwargs,252        )253 254 255__all__ = ["Qwen2VLProcessor"]256