Find video path for a video_id using on-demand lookup. Checks in-memory cache first, then disk cache, then searches 3 subdirectories. Dis

Find 2fps video path for a video_id using on-demand lookup. Checks in-memory cache first, then disk cache, then searches subdirectories. Disc

Load video path cache from disk.

Load an HF dataset, caching locally under VIDEO_DATA_HOME.

Save video path cache to disk.

Shard inputs on their sequence dimension, optionally adding padding if needed. .. important:: If using intra-document ma

Convert extracted point tracking data to standardized trajectory format. Args: points_data: List of dicts with 'time' and 'point

Extract a segment from the video that contains the moment. Args: moment_start (float): Start time of the moment in second

Format input timestamp as text

Get the context parallel sub-mesh associated with a ``DeviceMesh`` created from :func:`build_world_mesh()`. :param device_mesh: The worl

Get the class of the named dataset

Get the default maximum number of threads allowed.

Build ground truth target ids for a point point: (x, y) normalized coordinates between 0 and 1 patch_idx_arr: Mapping of patch x,y coordinate

Get directory to save the eval results

Check if the point (x, y) is within the region defined by the boolean mask. Parameters: - point (tuple of floats): x/y-coordinate of the

Check if point falls within the segmentation mask region. (Copied from olmo.eval.evaluators.py)

Join two paths. :param path1: The first path. :param path2: The second path. :returns: The joined result.

Load model state in-place from `dir` Works for any combination of sharded/unshared checkpoints and sharded/unshared model

Load video frames with decord

Normalize the str to lower case and make them float numbers if possible.

Get the peak GPU memory usage in MB across all ranks. Only rank 0 will get the final result.

Build a batch by interleaving the text tokens from tokenizing `message_list` and the multi-modal tokens from `multi_model_tokens`

Strip leading '#X' and capitalize.

Filter point_tracks to only include points AFTER initial_points Args: point_tracks: List of frame data with points

Filter frames_data to only include frames that match actual video timestamps. Uses numpy broadcasting for efficient comparison.

Filter frames_data to only include frames that match actual video timestamps. Uses numpy broadcasting for efficient comparison.

Return dir with annotations: v1 path (VIDEO_DIR) or v2 path (HOME).

Convert image points that are already sorted and scaled into a string

Format points for a single frame in video as {id: [x, y], ...} Shared by: - `_format_video_point_track_per_frame`

Return (zip_names, frames_root) for each split.

Get path for the preprocessed rows_with_clips pickle.

Load/cache annotations across all tasks/splits, return {(data_split, video_name): fps}. video_name is the video identifier without file exten

Load video list for a specific dataset (k600 or k700)

Extract tar files and decode mp4s to frames.

Look up the video path from the pre-built cache. Args: video_id: the `video` column value from HF (video_id, no extension)