msdnet.gpuoperations module¶

Module implementing network operations on GPU using Numba.

msdnet.gpuoperations.get1dgridsize(sz, tpb=1024)[source]¶

Return CUDA grid size for 1d arrays.

Parameters

sz – input array size
tpb – (optional) threads per block

msdnet.gpuoperations.get2dgridsize(sz, tpb=(8, 8))[source]¶

Return CUDA grid size for 2d arrays.

Parameters

sz – input array size
tpb – (optional) threads per block

class msdnet.gpuoperations.GPUImageData(shape, dl, nin)[source]¶

Bases: object

Object that represents a set of 2D images on GPU.

Parameters

shape – total shape of all images
dl – list of dilations in the network
nin – number of input images of network

set_block_size(imshape)[source]¶: Set CUDA grid sizes to be used.

setimages(ims)[source]¶

Set data to set of images.

Parameters: ims – set of images

setscalars(scl, start=0)[source]¶

Set each image to a scalar.

Parameters: scl – scalar values

fill(val, start=None, end=None)[source]¶

Set image data to single scalar value.

Parameters: val – scalar value

copy(start=None, end=None)[source]¶: Return copy of image data.

get(start=None, end=None)[source]¶: Return image data.

add(val, i)[source]¶

Add scalar to single image.

Parameters

val – scalar to add
i – index of image to add value to

mult(val, i)[source]¶

Multiply single image with value.

Parameters

val – value
i – index of image to multiply

prepare_forw_conv(f)[source]¶

Prepare for forward convolutions.

Parameters: f – convolution filters

forw_conv(i, outidx, dl)[source]¶

Perform forward convolutions

Parameters

i – image index to compute
outidx – image index to write output to
dl – dilation list

prepare_back_conv(f)[source]¶

Prepare for backward convolutions.

Parameters: f – convolution filters

back_conv(outidx, dl)[source]¶

Perform backward convolutions

Parameters

outidx – image index to write output to
dl – dilation list

relu(i)[source]¶: Apply ReLU to single image.

relu2(i, dat, j)[source]¶: Apply backpropagation ReLU to single image.

combine_all_all(dat, w)[source]¶: Compute linear combinations of images.

prepare_gradient()[source]¶: Prepare for gradient computation.

filtergradientfull(ims)[source]¶: Compute gradients for filters.

weightgradientall(delta)[source]¶: Compute gradients for weights.

sumall()[source]¶: Compute image sums.

softmax()[source]¶: Compute softmax.

msdnet.gpuoperations.fastmult_impl(a, b, out)[source]¶

msdnet.gpuoperations.fastsumall_impl(a, out)[source]¶