| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| import math |
|
|
| import numpy as np |
| import torch |
| from torchvision.transforms.v2 import functional as tvF |
|
|
| from transformers.image_processing_utils import BatchFeature |
| from transformers.image_utils import PILImageResampling, SizeDict |
| from transformers.processing_utils import Unpack, VideosKwargs |
| from transformers.utils import TensorType, auto_docstring, logging |
| from transformers.video_processing_utils import BaseVideoProcessor |
| from transformers.video_utils import ( |
| VideoMetadata, |
| group_videos_by_shape, |
| make_batched_metadata, |
| make_batched_videos, |
| reorder_videos, |
| ) |
|
|
|
|
| logger = logging.get_logger(__name__) |
|
|
|
|
| class Lfm3VlVideoProcessorInitKwargs(VideosKwargs, total=False): |
| r""" |
| patch_size (`int`, *optional*, defaults to 16): |
| The spatial patch size of the vision encoder. |
| temporal_patch_size (`int`, *optional*, defaults to 2): |
| The temporal patch size of the vision encoder. |
| merge_size (`int`, *optional*, defaults to 2): |
| The merge size of the vision encoder to llm encoder. |
| min_frames (`int`, *optional*, defaults to 4): |
| The minimum number of frames to sample from video. |
| max_frames (`int`, *optional*, defaults to 768): |
| The maximum number of frames to sample from video. |
| cap_pixels_per_frame (`bool`, *optional*): |
| Whether to cap each frame's pixel cost the way the reference implementation (qwen-vl-utils) does: |
| per-frame pixels are limited to `min(max_video_tokens * factor**2, size["longest_edge"] / num_frames)` |
| and floored at `1.05 * size["shortest_edge"]`, so token cost scales with clip duration. Without the |
| cap, videos that sample few frames spend the whole `size["longest_edge"]` budget on those frames and |
| keep near-native per-frame resolution, so a short clip can cost almost as many tokens as a long |
| video. If unset, the current uncapped behavior is kept and a warning is emitted: the default will |
| change to `True` in v5.22, after which the argument will be removed. |
| max_video_tokens (`int`, *optional*, defaults to 768): |
| The per-frame token ceiling applied by `cap_pixels_per_frame`, in vision tokens per frame |
| (qwen-vl-utils' `VIDEO_MAX_TOKEN_NUM`). |
| """ |
|
|
| patch_size: int |
| temporal_patch_size: int |
| merge_size: int |
| min_frames: int |
| max_frames: int |
| cap_pixels_per_frame: bool |
| max_video_tokens: int |
|
|
|
|
| def smart_resize( |
| num_frames: int, |
| height: int, |
| width: int, |
| temporal_factor: int = 2, |
| factor: int = 28, |
| min_pixels: int = 112 * 112, |
| max_pixels: int = 14 * 14 * 2 * 2 * 2 * 6144, |
| ): |
| if num_frames < temporal_factor: |
| raise ValueError(f"t:{num_frames} must be larger than temporal_factor:{temporal_factor}") |
| if height < factor or width < factor: |
| scale = max(factor / height, factor / width) |
| height = int(height * scale) |
| width = int(width * scale) |
|
|
| if max(height, width) / min(height, width) > 200: |
| raise ValueError( |
| f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}" |
| ) |
| h_bar = round(height / factor) * factor |
| w_bar = round(width / factor) * factor |
| t_bar = round(num_frames / temporal_factor) * temporal_factor |
|
|
| if t_bar * h_bar * w_bar > max_pixels: |
| beta = math.sqrt((num_frames * height * width) / max_pixels) |
| h_bar = max(factor, math.floor(height / beta / factor) * factor) |
| w_bar = max(factor, math.floor(width / beta / factor) * factor) |
| elif t_bar * h_bar * w_bar < min_pixels: |
| beta = math.sqrt(min_pixels / (num_frames * height * width)) |
| h_bar = math.ceil(height * beta / factor) * factor |
| w_bar = math.ceil(width * beta / factor) * factor |
|
|
| return h_bar, w_bar |
|
|
|
|
| def _check_video_fps(metadata: VideoMetadata) -> None: |
| if metadata.fps is not None and not (math.isfinite(metadata.fps) and metadata.fps > 0): |
| raise ValueError(f"`video_metadata.fps` must be a positive number, got {metadata.fps}") |
|
|
|
|
| def _check_video_metadata_count(videos, video_metadata) -> None: |
| |
| if video_metadata is None: |
| return |
| videos = make_batched_videos(videos) |
| entries = make_batched_metadata(videos, video_metadata) if video_metadata != [] else [] |
| if len(entries) != len(videos): |
| raise ValueError( |
| f"Received {len(videos)} videos but {len(entries)} `video_metadata` entries; pass one entry per video" |
| ) |
|
|
|
|
| @auto_docstring |
| class Lfm3VlVideoProcessor(BaseVideoProcessor): |
| resample = PILImageResampling.BICUBIC |
| size = {"shortest_edge": 4_096, "longest_edge": 25_165_824} |
| image_mean = [0.5, 0.5, 0.5] |
| image_std = [0.5, 0.5, 0.5] |
| do_resize = True |
| do_rescale = True |
| do_normalize = True |
| do_convert_rgb = True |
| patch_size = 16 |
| temporal_patch_size = 2 |
| merge_size = 2 |
| min_frames = 4 |
| max_frames = 768 |
| do_sample_frames = True |
| cap_pixels_per_frame = False |
| max_video_tokens = 768 |
| valid_kwargs = Lfm3VlVideoProcessorInitKwargs |
| model_input_names = ["pixel_values_videos", "video_grid_thw"] |
| max_image_size = {"longest_edge": 28 * 28 * 2 * 30000} |
| max_duration = None |
| num_frames = None |
| fps = 2.0 |
|
|
| def __init__(self, **kwargs: Unpack[Lfm3VlVideoProcessorInitKwargs]): |
| super().__init__(**kwargs) |
|
|
| def sample_frames( |
| self, metadata: VideoMetadata, num_frames: int | None = None, fps: int | float | None = None, **kwargs |
| ): |
| """Decoded frames whose metadata carries no fps are sampled and timestamped at the 24 fps the shared sampler |
| assumes (with a warning); pass `video_metadata` with the clip's fps for accurate timestamps.""" |
| _check_video_fps(metadata) |
| if fps is not None and not (math.isfinite(fps) and fps > 0): |
| raise ValueError(f"`fps` must be a positive number, got {fps}") |
| if fps is not None and num_frames is not None: |
| raise ValueError("`num_frames` and `fps` are mutually exclusive arguments, please use only one!") |
|
|
| total_num_frames = metadata.total_num_frames |
| fps = fps if fps is not None else self.fps |
|
|
| |
| if num_frames is None and fps is not None: |
| if metadata.fps is None: |
| metadata.fps = 24 |
| logger.warning_once( |
| "Asked to sample `fps` frames per second but no video metadata was provided which is required when sampling with `fps`. " |
| "Defaulting to `fps=24`. Please provide `video_metadata` for more accurate results." |
| ) |
| num_frames = int(total_num_frames / metadata.fps * fps) |
| num_frames = min(max(num_frames, self.min_frames), self.max_frames, total_num_frames) |
|
|
| if num_frames is None: |
| num_frames = min(max(total_num_frames, self.min_frames), self.max_frames) |
|
|
| indices = np.linspace(0, total_num_frames - 1, num_frames).round().astype(int) |
|
|
| return indices |
|
|
| def resize( |
| self, |
| videos: "torch.Tensor", |
| size: SizeDict, |
| resample: "PILImageResampling | tvF.InterpolationMode | int | None", |
| factor: int, |
| temporal_factor: int, |
| cap_pixels_per_frame: bool | None = None, |
| **kwargs, |
| ) -> "torch.Tensor": |
| """Resize dynamically based on input video aspect ratio.""" |
| if not size.shortest_edge or not size.longest_edge: |
| raise ValueError(f"`size` dict must contain 'shortest_edge' and 'longest_edge' keys but got {size}.") |
|
|
| num_frames = videos.shape[1] |
| max_pixels = size.longest_edge |
| if cap_pixels_per_frame: |
| |
| frame_cap = self.max_video_tokens * factor * factor |
| pixels_per_frame = max(min(frame_cap, size.longest_edge // num_frames), int(size.shortest_edge * 1.05)) |
| max_pixels = pixels_per_frame * num_frames |
|
|
| height, width = videos.shape[-2:] |
| resized_height, resized_width = smart_resize( |
| height=height, |
| width=width, |
| num_frames=num_frames, |
| factor=factor, |
| temporal_factor=temporal_factor, |
| min_pixels=size.shortest_edge, |
| max_pixels=max_pixels, |
| ) |
| return super().resize( |
| image=videos, |
| size=SizeDict(height=resized_height, width=resized_width), |
| resample=resample, |
| ) |
|
|
| def patchify( |
| self, |
| videos: "torch.Tensor", |
| patch_size: int, |
| merge_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 |
|
|
| |
| 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 // merge_size, |
| merge_size, |
| patch_size, |
| grid_w // merge_size, |
| merge_size, |
| patch_size, |
| ) |
| patches = patches.permute(0, 1, 4, 7, 5, 8, 3, 2, 6, 9) |
| flatten_patches = patches.reshape( |
| batch_size, |
| grid_t * grid_h * grid_w, |
| channel * temporal_patch_size * patch_size * patch_size, |
| ) |
|
|
| return flatten_patches, grid_t, grid_h, grid_w |
|
|
| def _preprocess( |
| self, |
| videos: list["torch.Tensor"], |
| do_resize: bool, |
| size: SizeDict, |
| 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, |
| patch_size: int | None = None, |
| temporal_patch_size: int | None = None, |
| merge_size: int | None = None, |
| cap_pixels_per_frame: bool | None = None, |
| return_tensors: str | TensorType | None = None, |
| **kwargs, |
| ): |
| if cap_pixels_per_frame is None: |
| logger.warning_once( |
| "Lfm3Vl video processing does not apply the per-frame pixel cap the reference " |
| "implementation (qwen-vl-utils) applies, so some videos cost far more tokens than they " |
| "would there. In v5.22 the capped behavior will become the default and " |
| "`cap_pixels_per_frame` will be removed. Pass `cap_pixels_per_frame=True` to adopt the " |
| "reference behavior now, or `False` to keep the current behavior and silence this " |
| "warning." |
| ) |
| cap_pixels_per_frame = False |
| |
| grouped_videos, grouped_videos_index = group_videos_by_shape(videos) |
| resized_videos_grouped = {} |
| for shape, stacked_videos in grouped_videos.items(): |
| if do_resize: |
| stacked_videos = self.resize( |
| videos=stacked_videos, |
| size=size, |
| resample=resample, |
| factor=patch_size * merge_size, |
| temporal_factor=temporal_patch_size, |
| cap_pixels_per_frame=cap_pixels_per_frame, |
| ) |
| resized_videos_grouped[shape] = stacked_videos |
| resized_videos = reorder_videos(resized_videos_grouped, grouped_videos_index) |
|
|
| |
| |
| grouped_videos, grouped_videos_index = group_videos_by_shape(resized_videos) |
| processed_videos_grouped = {} |
| processed_grids = {} |
| for shape, stacked_videos in grouped_videos.items(): |
| |
| 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, |
| merge_size=merge_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, |
| ) |
|
|
| def get_num_of_video_patches(self, num_frames: int, height: int, width: int, videos_kwargs=None): |
| """ |
| A utility that returns number of video patches a given video size. |
| |
| Args: |
| num_frames (`int`): |
| Number of frames in the input video. |
| height (`int`): |
| Height of the input video. |
| width (`int`): |
| Width of the input video. |
| videos_kwargs (`dict`, *optional*) |
| Any kwargs to override defaults of the video processor. |
| Returns: |
| `int`: Number of video patches per video. |
| """ |
| videos_kwargs = videos_kwargs if videos_kwargs is not None else {} |
| size = videos_kwargs.get("size", None) or self.size |
| min_pixels = size["shortest_edge"] |
| max_pixels = size["longest_edge"] |
| patch_size = videos_kwargs.get("patch_size", None) or self.patch_size |
| merge_size = videos_kwargs.get("merge_size", None) or self.merge_size |
| temporal_patch_size = videos_kwargs.get("temporal_patch_size", None) or self.temporal_patch_size |
| cap_pixels_per_frame = videos_kwargs.get("cap_pixels_per_frame", None) or self.cap_pixels_per_frame |
|
|
| factor = patch_size * merge_size |
| if cap_pixels_per_frame: |
| |
| frame_cap = self.max_video_tokens * factor * factor |
| pixels_per_frame = max(min(frame_cap, max_pixels // num_frames), int(min_pixels * 1.05)) |
| max_pixels = pixels_per_frame * num_frames |
|
|
| resized_height, resized_width = smart_resize( |
| num_frames, |
| height, |
| width, |
| temporal_factor=temporal_patch_size, |
| factor=factor, |
| min_pixels=min_pixels, |
| max_pixels=max_pixels, |
| ) |
| grid_h, grid_w = resized_height // patch_size, resized_width // patch_size |
| grid_t = (num_frames + -num_frames % temporal_patch_size) // temporal_patch_size |
| return grid_t * grid_h * grid_w |
|
|
| def preprocess(self, videos, **kwargs): |
| _check_video_metadata_count(videos, kwargs.get("video_metadata")) |
| |
| |
| |
| num_frames, fps = kwargs.get("num_frames"), kwargs.get("fps") |
| if num_frames is not None and fps is not None: |
| raise ValueError("`num_frames` and `fps` are mutually exclusive arguments, please use only one!") |
| if num_frames is not None or (fps is None and self.num_frames is not None): |
| kwargs["fps"] = None |
| elif fps is not None: |
| kwargs["num_frames"] = None |
| return super().preprocess(videos, **kwargs) |
|
|
|
|
| __all__ = ["Lfm3VlVideoProcessor"] |
|
|
