from typing import List, Optional, Union import numpy as np from transformers import BatchFeature from transformers.tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput from transformers.image_utils import ImageInput, VideoInput from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack, VideosKwargs class HithinkOmniVideosProcessorKwargs(VideosKwargs, total=False): fps: Union[List[float], float] class HithinkOmniProcessorKwargs(ProcessingKwargs, total=False): videos_kwargs: HithinkOmniVideosProcessorKwargs _defaults = { "text_kwargs": { "padding": False, }, "videos_kwargs": {"fps": 2.0}, } class HithinkOmniProcessor(ProcessorMixin): r""" Constructs a HithinkOmni processor which wraps a Qwen2.5-VL image processor and a HithinkOmni tokenizer into a single processor. [`HithinkOmniProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`PreTrainedTokenizerFast`]. See the [`~HithinkOmniProcessor.__call__`] and [`~HithinkOmniProcessor.decode`] for more information. Args: image_processor ([`Qwen2VLImageProcessor`], *optional*): The image processor is a required input. feature_extractor ([`WhisperFeatureExtractor`], *optional*): The feature extractor is a required input. tokenizer ([`PreTrainedTokenizerFast`], *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", "feature_extractor", "tokenizer"] valid_kwargs = ["chat_template"] image_processor_class = "Qwen2VLImageProcessor" feature_extractor_class = "WhisperFeatureExtractor" tokenizer_class = "PreTrainedTokenizerFast" def __init__(self, image_processor=None, feature_extractor=None, tokenizer=None, chat_template=None, **kwargs): tokenizer.model_input_names = ["input_ids", "attention_mask"] # do not include token_type_ids super().__init__(image_processor, feature_extractor, tokenizer, chat_template=chat_template) self.image_token = getattr(tokenizer, 'image_token', '<|image_pad|>') self.video_token = getattr(tokenizer, 'video_token', '<|video_pad|>') self.chat_template = tokenizer.chat_template if chat_template is None else chat_template def __call__( self, images: ImageInput = None, text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None, videos: VideoInput = None, audios: Union[np.ndarray, List[np.ndarray]] = None, sampling_rate: Optional[int] = None, **kwargs: Unpack[HithinkOmniProcessorKwargs], ) -> 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 PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__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 Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`. Args: 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. 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). 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. audios (`np.ndarray`, `List[np.ndarray]`): The audio or batch of audios to be prepared. Each audio can be a NumPy array. sampling_rate (`int`, defaults to 16000): The sampling rate at which the audio files should be digitalized expressed in hertz (Hz). 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( HithinkOmniProcessorKwargs, tokenizer_init_kwargs=self.tokenizer.init_kwargs, **kwargs, ) if images is not None: image_inputs = self.image_processor(images=images, videos=None, **output_kwargs["images_kwargs"]) image_grid_thw = image_inputs["image_grid_thw"] else: image_inputs = {} image_grid_thw = None if videos is not None: videos_inputs = self.image_processor(images=None, videos=videos, **output_kwargs["images_kwargs"]) video_grid_thw = videos_inputs["video_grid_thw"] fps = output_kwargs["videos_kwargs"].pop("fps", 2.0) if isinstance(fps, (int, float)): second_per_grid_ts = [self.image_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.image_processor.temporal_patch_size / tmp for tmp in fps] else: raise ValueError( f"The length of fps ({len(fps) if hasattr(fps, '__len__') else fps}) must be equal to the length of video_grid_thw ({len(video_grid_thw)}) or fps should be a single number." ) videos_inputs.update({"second_per_grid_ts": second_per_grid_ts}) else: videos_inputs = {} video_grid_thw = None if not isinstance(text, list): text = [text] if image_grid_thw 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]: text[i] = text[i].replace( self.image_token, "<|placeholder|>" * (image_grid_thw[index].prod() // merge_length), 1, ) index += 1 text[i] = text[i].replace("<|placeholder|>", self.image_token) if video_grid_thw is not None: merge_length = self.image_processor.merge_size**2 index = 0 for i in range(len(text)): while self.video_token in text[i]: text[i] = text[i].replace( self.video_token, "<|placeholder|>" * (video_grid_thw[index].prod() // merge_length), 1, ) index += 1 text[i] = text[i].replace("<|placeholder|>", self.video_token) if audios is not None: audio_inputs = self.feature_extractor( audios, sampling_rate=sampling_rate, return_attention_mask=True, padding="max_length", **kwargs ) audio_inputs["feature_attention_mask"] = audio_inputs.pop( "attention_mask" ) # rename attention_mask to prevent conflicts later on audio_output_lengths = self.get_feat_extract_output_lengths( audio_inputs['feature_attention_mask'].sum(-1) ) index = 0 for i in range(len(text)): while "<|AUDIO|>" in text[i]: text[i] = text[i].replace( "<|AUDIO|>", "<|placeholder|>" * audio_output_lengths[index], 1 ) index += 1 text[i] = text[i].replace("<|placeholder|>", "<|AUDIO|>") else: audio_inputs = {} text_inputs =self.tokenizer(text, **output_kwargs["text_kwargs"]) return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs, **audio_inputs}) @staticmethod def get_feat_extract_input_length(audio_length): """ Computes the input length of the audio encoder (i.e. output of the feature extractor) e.g. 30-second audio has 480,000 samples (sampling_rate = 16,000), the feature length will be 3,000 """ return int(np.ceil((audio_length - 40) / 160)) # 第一帧需要200样本，后续每帧需要160样本 @staticmethod def get_feat_extract_output_lengths(input_lengths): """ Computes the output length of the convolutional layers and the output length of the audio encoder """ input_lengths = (input_lengths - 1) // 2 + 1 output_lengths = (input_lengths - 2) // 2 + 1 return output_lengths def featurize_audio_chunk(self, audio: np.ndarray, is_last: bool, n_extracted_frames: int = 0, **kwargs): """ Extract the features from the audio chunk during streaming inference """ n_frames = (len(audio) - 40) / 160 # 第一帧需要200样本，后续每帧需要160样本 n_frames = int(np.ceil(n_frames) if is_last else np.floor(n_frames)) n_new_frames = n_frames - n_extracted_frames i_end = n_frames * 160 + 40 i_start = max(0, (n_extracted_frames + 1 - 3) * 160) # 滑窗需要400样本，即最少3帧 if n_new_frames <= 0 or n_frames < 2: return a = audio[i_start: i_end] # 截取计算new frames需要的chunk if is_last and (n_pad := int(np.ceil(len(a) / 160)) * 160 - len(a)): # pad to multiple of 160 a = np.pad(a, [0, n_pad]) features = self.feature_extractor( a, sampling_rate=self.feature_extractor.sampling_rate, padding='do_not_pad', **kwargs )['input_features'] return features[:, :, -n_new_frames:] def batch_decode(self, *args, **kwargs): """ This method forwards all its arguments to PreTrainedTokenizerFast'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 PreTrainedTokenizerFast'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): """ 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,)`. Returns: `List[str]`: The decoded text. """ return self.tokenizer.batch_decode( generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False ) @property def model_input_names(self): tokenizer_input_names = self.tokenizer.model_input_names image_processor_input_names = self.image_processor.model_input_names feature_extractor_input_names = self.feature_extractor.model_input_names return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names + feature_extractor_input_names + ["feature_attention_mask"])) # audio