The internal method called by the hooked layers after they are run.

Internal method to remove a hook, restoring the original forward method.

Generates an upsampled activation sample. Params: target_shape Shape of the output array. source_shape The centered shape of

Makes a visualiztion image of activation data overlaid on the image. Params: image The original image. data The single channe

Finds the smallest subrectangle containing all the nonzeros of mask, matching the aspect ratio of shape, and where the zoom-up ratio is no mor

Returns (r, b, l, r) boundaries so that all nonzero pixels in mask have locations (i, j) with t <= i < b, and l <= j < r.

Input: set of rows with 'number' fields (or another field name key). Output: array such that a[number] = the row with the given number.

Returns label_counts, category_activation_counts, and intersection_counts, across the data set, counting the pixels of intersection between u

Returns label_mean, label_variance, unit_mean, unit_variance, and cross_covariance across the data set. label_mean, label_variance (inde

Returns iqr and iqr_level. This is a performance-sensitive function. Best performance is achieved with a counting scheme which assumes

Returns maxiou and maxiou_level across the data set, one per layer. This is a performance-sensitive function. Best performance is achie

Collects (estimated) quantile information and (exact) sorted top-K lists for every channel in the retained layers of the model. Returns

3x3 convolution with padding

3x3 convolution with padding

Translate output sample indices (small numbers indexing the sample) to input sample indices (larger number indexing the original full

Composes a list of (k, s, p) into a single total scale and offset.

Converts a list of modules to (kernel_size, stride, padding)

Overrides the forward method of the given layer, to replace the output with a convex combination of the output and the given value, as sp

Creates 'ablation' and 'replacement' properties on the model. The output of each layer is replaced by the convex combination of the actua