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

Evaluate caption using OpenAI API. :param prediction: predicted caption :param data: ground truth data :param openai_api_key: OpenAI

Evaluate a multiple choice instance.

Evaluate an open question instance

Evaluate predictions for a single frame. Args: pred_points: List of predicted (x, y) coordinates gt_points: List of grou

Evaluate predictions for a single video. Args: pred_tracks: Predicted trajectory data Format: [{

Build HTML visualizations for an examples ex: The example (after preprocessing) to show preprocessor: The preprocessor used to preprocessor t

Extract a clip from source_path[start:end] -> output_path using ffmpeg. Returns (success, message).

Extract a clip from a video using ffmpeg. Args: video_path (str): Path to the input video start_time (float): Start time

Extract all frames from a video to JPEG files using ffmpeg.

Extract points from MolmoPoint image output text return points: [n_points, 4] array of (object_id, image_num, x, y) points

Exact all forms of numbers from a string with regex.

Extract the yes/no predction from the original video llm output

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

Extract points from video_point_track_per_frame prompt style. Text format example: time 0.50 {0: [34.0, 63.0], 1: [72.0, 49.0

Extract points from video_point_track_all_frames_with_occlusion format. Text format example (yes if occluded): time 0.50 {0:

Extract points from MolmoPoint video output text return points: [n_points, 4] array of (object_id, timestamp, x, y) points

Extract points from video_point_ground_start_end format prompt style. Text format example: 0: ([34.0, 63.0, 0.50], [35.0, 64.0, 1.00]

Compute F1 score from precision and recall. (Copied from olmo.eval.evaluators.py)

Convert a single or a list of urls into the corresponding `np.array` objects. If a single url is passed, the return value will be a

Find the first padding frame for each batch. Args: pooled_patches_idx (torch.Tensor): Tensor of shape (B, NP, G) indicating the pooled pa

Find the source video file, trying known extensions.

Flatten a multi-dimensional ``DeviceMesh`` into a 1D ``DeviceMesh``. :param device_mesh: The multi-dimensional ``DeviceMesh`` to flatten.

Applies system formatting to ith message from a sequence of messages

Applies system formatting to ith message from a sequence of messages

Format multi-image points

Format video points for counting objects across frames.

Format multi-image pointing and counting. Generates questions and answers for pointing to and counting objects across multiple images.

Legacy method for formatting initial input points for user prompt.

Format video points for tracking objects across frames. Keep only frames that match actual video timestamps. Sample frames at

Format video points for tracking objects across frames. Keep only frames that match actual video timestamps. Sample frames at

Format multi-image pointing and counting. Generates questions and answers for pointing to and counting objects across multiple images.

Run a query through an OpenAI model

Run a query through an OpenAI model

Extract subtitles that overlap with the given time range.

Get the default device.

Get the data parallel sub-mesh associated from a ``DeviceMesh`` created by :func:`build_world_mesh()`. .. important:: This is the me

Get the right sub-mesh for a data parallel model wrapper like FSDP or DDP from a ``DeviceMesh`` created by :func:`build_world_mesh()`. .

Get the size of a local or remote file in bytes. :param path: Path/URL to the file.

Get the font size based on the size of the image.

Transform coordinates from original image to resized image with padding. Args: x_orig: X coordinate in original image y_

Get a FSDP1 wrap policy for this model.