Translates an array in-place by the shift xy, wrapping at the edges.

Sets y = a*x + y for float a and float32 arrays x, y; returns y.

Rolls the image, feature maps.

Makes dataclass-based message types for IPC.

Returns a step's parameter update given a loss/gradient evaluation function.

Returns the dot product of two float32 arrays with the same shape.

Returns the layer names of the network.

Returns the scale factor vs. the image and the number of channels.

Sets the current model input to a PIL image.

Returns 1/2 the L2 norm squared.

Resamples an image array in CxHxW format to a new HxW size. The interpolation is performed in floating point and the result dtype is numpy.float32

Resizes image to fit into a size-by-size square.

Rolls a 3D array in-place by a shift of one element. Axes 1 and 2 only.

Consume the iterable.

Normalizes an array in-place to have an L1 norm equal to its size.

Parses a list of name:number pairs into a normalized dict of weights.

Performs preprocessing tasks on the input images.

Rolls an individual set of feature maps in-place.

This method runs in the new process.

Checks for abnormal pool process termination.

Computes the Gram matrix corresponding to a feature map.

Returns the pth power of the p-norm and its gradient.

Preprocesses a PIL image into params format.

Averages the set of feature maps for an image over multiple passes to obscure tiling.

Enqueues a request.

Sets the worker which will process the next request to worker 0.

Returns the sum of the absolute values of the given float32 array.

Propagates feature maps and Gram matrices to all TileWorkers.

Sets the maximum number of MKL threads for this process.

Returns the first argument if the terminal has a light background, the second if it has a dark background, and the third if it cannot determine.