sleap.nn.utils

sleap.nn.utils#

This module contains generic utilities used for training and inference.

sleap.nn.utils.compute_iou(bbox1: ndarray, bbox2: ndarray) → float[source]#

Computes the intersection over union for a pair of bounding boxes.

Parameters:

  • bbox1 – Bounding box specified by corner coordinates [y1, x1, y2, x2].

  • bbox2 – Bounding box specified by corner coordinates [y1, x1, y2, x2].

Returns:

A float scalar calculated as the ratio between the areas of the intersection and the union of the two bounding boxes.

sleap.nn.utils.group_array(X: ndarray, groups: ndarray, axis: int = 0) → Dict[ndarray, ndarray][source]#

Groups an array into a dictionary keyed by a grouping vector.

Parameters:

  • X – Numpy array with length n along the specified axis.

  • groups – Vector of n values denoting the group that each slice of X should be assigned to. This is also referred to as an indicator, indexing, class, or labels vector.

  • axis – Dimension of X to group on. The length of this axis in X must correspond to the length of groups.

Returns:

A dictionary with keys mapping each unique value in groups to a subset of X.

References

See this blog post for performance comparisons of different approaches.

Example:

>>> group_array(np.arange(5), np.array([1, 5, 2, 1, 5]))
{1: array([0, 3]), 5: array([1, 4]), 2: array([2])}
sleap.nn.utils.match_points(points1: Tensor, points2: Tensor) → Tuple[Tensor, Tensor][source]#

Match closest points across two sets.

Parameters:

  • points1 – A set of reference points of shape (N, D).

  • points2 – A set of unmatched points of shape (M, D).

Returns:

A tuple of (inds1, inds2) with the indices of matching points across the two lists of points.

inds1 is a tf.int32 tensor of shape (n,) where n <= N that indexes into points1.

inds2 is a tf.int32 tensor of shape (m,) where m <= M that indexes into points2.

Notes

This uses Euclidean distance and the Hungarian algorithm for matching.

See also: tf_linear_sum_assignment

sleap.nn.utils.reset_input_layer(keras_model: Model, new_shape: Tuple[int | None, int | None, int | None, int] | None = None)[source]#

Returns a copy of keras_model with input shape reset to new_shape.

This method was modified from https://stackoverflow.com/a/58485055.

Parameters:

  • keras_model – tf.keras.Model to return a copy of (with input shape reset).

  • new_shape – Shape of the returned model’s input layer.

Returns:

A copy of keras_model with input shape new_shape.

sleap.nn.utils.tf_linear_sum_assignment(cost_matrix: Tensor) → Tensor#

Run linear_sum_assignment as a TensorFlow function.

Parameters:

cost_matrix – Cost matrix of shape (n_src, n_dst). Make sure to replace NaN`s with `np.inf.

Returns:

A tuple of (row, col) with the indices of the optimal assignments.

Contents