Parse point_ground_start_end predictions to extract grounded objects with start points. Used for creating synthetic single_point_track_per_fr

Assumes: - line 1 = question - lines 2..n = options (one per line) Accepts options like "(A) foo", "A) foo", "A.

Parse temporal segments from text output. Expected format: "[start1-end1], [start2-end2], ..." Returns a list of [start, end]

Parse single_point_track_per_frame model output. Expected format matches data_formatter.py: "[x, y, time], [x, y, time], [x, y, time

Extracts point coordinates and timestamps from raw text predictions and calculates frame indices based on video FPS. Converts into s

Convert trajectory data to HOTA format. Returns: data: dict with keys needed for HOTA computation

Process all clips for a single video in parallel. Args: video_data (tuple): Tuple of (video_path, clips_list) output_pat

Score tokens based on results from `TemporalTokenScorer`, then sort them by topk to prune them. Args: x (torch.Tensor): The input

Randomly prune tokens for sanity check. Args: valid_mask (torch.Tensor): A mask indicating valid tokens. num_to_prune (torch

Call `resource_path` with the given args for the rank0 process, other ranks will wait until rank0 is done downloading. Should be called

Calculates relaxed correctness. The correctness tolerates certain error ratio defined by max_relative_change. See https://arxiv.org/pdf/2203.

Resize an image while padding to preserve uts aspect ratio.

Returns the beams with the highest perturbed log probabilities.

Save model and optimizer state dictionaries. The model state can be a sharded model, in which case this method will correctly handle the opti

Save model, optim, and other training state to a local or remote directory unsharded :warning This can be very slow if saving to a remote dir

Score the final predictions found by beam search. Returns a tensor of the final sequence scores of shape `(batch_size, beam_size)`.

Whether the checkpointed dataloader already produced example number `on_example`

Vectorized 2D knapsack dynamic program solver

Divide in image of size [w, h] in up to max_num_patches of size patch_size

:param rank0_only: INFO and below messages will only be emitted on the rank0 process.

Split video using ffmpeg with re-encoding for reliability. Why re-encode instead of copy? - Codec copy (-c copy) cuts at non-keyframe positio

Table as listoflists to evenly spaces string

Template a multiple choice questions

Convert patch/location ids generated by the model to pixel coordinates text: str string output for the model target_ids: [n_points, 2

Handle multi-annotation data where we have many annotations for one multi-modal input

Handle multi-annotation data where we have many annotations for one multi-modal input

Normalizes the image to zero mean and unit variance.

Update the legacy config settings whose schemas have undergone backwards-incompatible changes.