lisbet.inference.embedding#

Embedding extraction for LISBET.

Functions

compute_embeddings(model_path, weights_path, ...)

Compute LISBET embeddings for every record in a dataset.
lisbet.inference.embedding.compute_embeddings(model_path, weights_path, data_format, data_path, data_scale=None, data_filter=None, window_size=200, window_offset=0, fps_scaling=1.0, batch_size=128, output_path=None, select_coords=None, rename_coords=None)[source]#

Compute LISBET embeddings for every record in a dataset.

This function loads an embedding model and processes an entire dataset, computing embeddings for each sequence.

Parameters:

  • model_path (str) – Path to the model config (JSON format).

  • weights_path (str) – Path to the HDF5 file containing the model weights.

  • data_format (str) – Format of the dataset to analyze.

  • data_path (str) – Path to the directory containing the dataset files.

  • data_scale (str | None) – Scaling string or None for auto-scaling.

  • data_filter (str | None) – Filter to apply when loading records.

  • window_size (int) – Size of the sliding window to apply on the input sequences.

  • window_offset (int) – Sliding window offset.

  • fps_scaling (float) – FPS scaling factor.

  • batch_size (int) – Batch size for inference.

  • output_path (str | None) – If given, embeddings will be saved as CSV files in this directory.

  • select_coords (str | None) – Optional subset string in the format ‘INDIVIDUALS;AXES;KEYPOINTS’, where each field is a comma-separated list or ‘*’ for all. If None, all data is loaded.

  • rename_coords (str | None) – Optional coordinate names remapping in the format ‘INDIVIDUALS;AXES;KEYPOINTS’, where each field is a comma-separated list of maps ‘old_id:new_id’ or ‘*’ for no remapping at that level. If None, original dataset names are used.

Returns:

A list of (sequence ID, embedding) tuples for each sequence.

Return type:

list[tuple[str, ndarray]]

Raises:

ValueError – If the loaded model is not an embedding model.