Decompose an image into a set of overlapping crops :return crop_arr: [n_crops, h, w, 3] The crops :return patch_idx: [overlap_patch_h, overla

Decompose an image into a set of overlapping crops :return crop_arr: [n_crops, h, w, 3] The crops :return mask_arr: [n_crops, h, w] T

Build ground truth target ids for an image

Build ground truth target ids for set of images

Build ground truth target ids for a video

Build a preprocessor that converts 'raw' image/text data from various tasks into tensors inputs/targets that can be passed to the mod

Build a :class:`~torch.distributed.device_mesh.DeviceMesh` suitable for the given parallel strategies. .. seealso:: Pass the mesh cr

Check if the given string a number.

Compute HOTA metrics given prepared data. Adapted from TrackEval HOTA implementation.

Pre-compute image shard boundaries for each GPU rank. Args: cum_image_bounds: Cumulative image counts for each input cum

Compute IoU (Intersection over Union) between two temporal segments.

Compute precision, recall, and F1 score for predicted segments. Uses IoU threshold to determine if a prediction matches a ground trut

Compute precision: correctly placed points / total predicted points. (Adapted from olmo.eval.evaluators.py)

Compute recall: correctly placed points / total ground truth points. (Adapted from olmo.eval.evaluators.py)

Computes TAP-Vid metrics (Jaccard, Pts. Within Thresh, Occ. Acc.) See the TAP-Vid paper for details on the metric computation. All inputs are

Convert config to HF-compatible config

Convert config to HF-compatible config

Convert model to HF-compatible model

Convert model to HF-compatible model

Creates a video file from a sequence of frames in a directory. Args: frames_dir (str): Directory containing the frames

Heuristic to detect format of the prediction text.

Call a function ``f`` in a distributed context from at most ``n`` ranks at a time. All ranks will eventually call the given function exactly

Download and create videos for specified sources. Args: dataset: 'train', 'eval', 'molmopoint-trackany', or 'molmopoint-tracksyn'

Prune tokens based on their validity and similarity scores. Easy version that only looks at the immediate previous frame. Args: valid

Parse the prediction from the generated response. Return the predicted index e.g., A, B, C, D. https://github.com/MMMU-Benchmark/MMMU/blo

Convert a directory of image frames to an H.264 .mp4 using ffmpeg. Frames are first subsampled for FPS conversion (if native_fps differs from

: param images: (batch_size, num_crops, num_patch, n_pixels)

: param images: (batch_size, num_crops, num_patch, n_pixels)

: param images: (batch_size, num_crops, num_patch, n_pixels)

Modify ``x`` in place to replace ``float("-inf")`` with the minimum value of the dtype when ``check_neg_inf`` is ``True`` and to replace ``fl