# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/muse_glimmer/modular_muse_glimmer.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_muse_glimmer.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # Copyright 2026 the HuggingFace 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 # limitations under the License. import itertools import math import torch from torchvision.transforms.v2 import functional as tvF from ...image_processing_utils import BatchFeature from ...image_utils import PILImageResampling, SizeDict from ...processing_utils import Unpack, VideosKwargs from ...utils import TensorType, auto_docstring, logging from ...utils.constants import IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD from ...video_processing_utils import BaseVideoProcessor from ...video_utils import VideoMetadata, group_videos_by_shape, reorder_videos logger = logging.get_logger(__name__) class MuseGlimmerVideoProcessorInitKwargs(VideosKwargs, total=False): """ patch_size (`int`, *optional*): The spatial patch size of the vision encoder, in pixels. temporal_patch_size (`int`, *optional*): The temporal patch size of the vision encoder, in frames. max_video_frame_tokens (`int`, *optional*): Maximum number of vision tokens per video frame; frames are resized to stay under this cap. merge_size (`int`, *optional*): Factor by which the patch grid is downsampled by pixel shuffling after the vision encoder. """ patch_size: int temporal_patch_size: int max_video_frame_tokens: int merge_size: int def smart_resize( height: int, width: int, patch_size: int, max_tokens: int, ) -> tuple[int, int]: """Pick the integer patch grid closest to the input aspect ratio under the token cap. Returns the resize target ``(target_height, target_width)`` in pixels. """ ideal_patches_height = height / patch_size ideal_patches_width = width / patch_size ratio = ideal_patches_width / ideal_patches_height if ideal_patches_height > 0 else 1.0 if ideal_patches_height * ideal_patches_width > max_tokens: ideal_patches_height = (max_tokens / ratio) ** 0.5 ideal_patches_width = ideal_patches_height * ratio candidates = list( set( itertools.product( [math.floor(ideal_patches_height), math.ceil(ideal_patches_height)], [math.floor(ideal_patches_width), math.ceil(ideal_patches_width)], ) ) ) candidates = [ (patches_height, patches_width) for patches_height, patches_width in candidates if patches_height >= 1 and patches_width >= 1 and patches_height * patches_width <= max_tokens ] if not candidates: candidates = [(max(1, round(ideal_patches_height)), max(1, round(ideal_patches_width)))] patches_height, patches_width = min(candidates, key=lambda grid: abs(grid[0] / grid[1] - height / width)) return patches_height * patch_size, patches_width * patch_size @auto_docstring class MuseGlimmerVideoProcessor(BaseVideoProcessor): resample = PILImageResampling.LANCZOS image_mean = IMAGENET_STANDARD_MEAN image_std = IMAGENET_STANDARD_STD default_to_square = True do_convert_rgb = True do_resize = True do_rescale = True do_normalize = True patch_size = 14 temporal_patch_size = 2 merge_size = 2 max_video_frame_tokens = 144 num_frames = 96 fps = 2.0 do_sample_frames = True valid_kwargs = MuseGlimmerVideoProcessorInitKwargs model_input_names = ["pixel_values_videos", "video_grid_thw"] def __init__(self, **kwargs: Unpack[MuseGlimmerVideoProcessorInitKwargs]): super().__init__(**kwargs) def _validate_preprocess_kwargs(self, **kwargs): # MuseGlimmer uses aspect_ratio_preserving_resize driven by patch_size, # not the standard `size` parameter. Temporarily disable do_resize so # the base validation doesn't raise an error kwargs["do_resize"] = False super()._validate_preprocess_kwargs(**kwargs) def resize( self, videos: torch.Tensor, resample: PILImageResampling | tvF.InterpolationMode | int | None, patch_size: int, merge_size: int, max_tokens: int, **kwargs, ) -> torch.Tensor: """Resize dynamically based on input video aspect ratio.""" height, width = videos.shape[-2:] resized_height, resized_width = smart_resize( height=height, width=width, patch_size=patch_size * merge_size, max_tokens=max_tokens, ) return super().resize( videos, size=SizeDict(height=resized_height, width=resized_width), resample=resample, antialias=True, ) def patchify( self, videos: torch.Tensor, patch_size: int, temporal_patch_size: int, ) -> tuple[torch.Tensor, int, int]: "Patchifies each video into flat layout of shape (`seq_len`, `patch_dim`) so we can concat dynamically shaped pixels." batch_size, num_frames, channel, resized_height, resized_width = videos.shape # Check that videos have `num_frames` divisible by `temporal_patch_size` if pad := -num_frames % temporal_patch_size: repeats = videos[:, -1:].expand(-1, pad, -1, -1, -1) videos = torch.cat((videos, repeats), dim=1) num_frames += pad grid_t = num_frames // temporal_patch_size grid_h, grid_w = resized_height // patch_size, resized_width // patch_size patches = videos.view( batch_size, grid_t, temporal_patch_size, channel, grid_h, patch_size, grid_w, patch_size, ) # Unlike Glm4v, each flattened patch is laid out (temporal, channel), not (channel, temporal). patches = patches.permute(0, 1, 4, 6, 2, 3, 5, 7) flatten_patches = patches.reshape( batch_size, grid_t * grid_h * grid_w, temporal_patch_size * channel * patch_size * patch_size, ) return flatten_patches, grid_t, grid_h, grid_w def sample_frames( self, metadata: VideoMetadata, temporal_patch_size: int | None = None, num_frames: int | None = None, fps: int | float | None = None, **kwargs, ): """ Default sampling function which uniformly samples the desired number of frames between 0 and total number of frames. If `fps` is passed along with metadata, `fps` frames per second are sampled uniformty. Arguments `num_frames` and `fps` are mutually exclusive. Args: metadata (`VideoMetadata`): Metadata of the video containing information about total duration, fps and total number of frames. temporal_patch_size (`int`, *optional*): The temporal patch size of the vision encoder. Number of sampled frames will be rounded to be divisible by frame factor. num_frames (`int`, *optional*): Maximum number of frames to sample. Defaults to `self.num_frames`. fps (`int` or `float`, *optional*): Target frames to sample per second. Defaults to `self.fps`. Returns: np.ndarray: Indices to sample video frames. """ if metadata.fps is None: logger.warning_once( "The `fps` of the input video could not be inferred. Defaulting to `fps=24`. " "Provide `video_metadata` for more accurate frame sampling." ) metadata.fps = 24 total_num_frames = metadata.total_num_frames num_frames = min(int(total_num_frames * fps / metadata.fps), num_frames, total_num_frames) num_frames = max(temporal_patch_size, (num_frames // temporal_patch_size) * temporal_patch_size) num_frames = min(num_frames, total_num_frames) indices = torch.linspace(0, total_num_frames - 1, num_frames).long() return indices def _preprocess( self, videos: list[torch.Tensor], do_resize: bool, do_convert_rgb: bool, resample: PILImageResampling | tvF.InterpolationMode | int | None, do_rescale: bool, rescale_factor: float, do_normalize: bool, image_mean: float | list[float] | None, image_std: float | list[float] | None, return_tensors: str | TensorType | None, patch_size: int, temporal_patch_size: int, max_video_frame_tokens: int, merge_size: int, disable_grouping: bool = False, **kwargs, ) -> BatchFeature: # Group videos by size for batched resizing grouped_videos, grouped_videos_index = group_videos_by_shape(videos) resized_videos_grouped = {} for shape, stacked_videos in grouped_videos.items(): if do_convert_rgb: stacked_videos = self.convert_to_rgb(stacked_videos) if do_resize: stacked_videos = self.resize( stacked_videos, patch_size=patch_size, merge_size=merge_size, max_tokens=max_video_frame_tokens, resample=resample, ) resized_videos_grouped[shape] = stacked_videos resized_videos = reorder_videos(resized_videos_grouped, grouped_videos_index) # Group videos by size for further processing # Needed in case do_resize is False, or resize returns videos with different sizes grouped_videos, grouped_videos_index = group_videos_by_shape(resized_videos) processed_videos_grouped = {} processed_grids = {} for shape, stacked_videos in grouped_videos.items(): # Fused rescale and normalize stacked_videos = self.rescale_and_normalize( stacked_videos, do_rescale, rescale_factor, do_normalize, image_mean, image_std ) patches, grid_t, grid_h, grid_w = self.patchify( stacked_videos, patch_size=patch_size, temporal_patch_size=temporal_patch_size, ) processed_videos_grouped[shape] = patches processed_grids[shape] = [[grid_t, grid_h, grid_w]] * len(stacked_videos) processed_videos = reorder_videos(processed_videos_grouped, grouped_videos_index) processed_grids = reorder_videos(processed_grids, grouped_videos_index) pixel_values_videos = torch.cat(processed_videos, dim=0) video_grid_thw = torch.tensor(processed_grids) return BatchFeature( data={"pixel_values_videos": pixel_values_videos, "video_grid_thw": video_grid_thw}, tensor_type=return_tensors, ) __all__ = ["MuseGlimmerVideoProcessor"]