msdnet.network module

Module implementing neural networks.

class msdnet.network.Network

Bases: abc.ABC

Base class for a neural network.

abstract forward(im, returnoutput=True)

Compute a forward pass of the network.

Parameters

• im – input image (channels x rows x columns)

• returnoutput – whether to return the output image (default: True)

Returns

output image (channels x rows x columns)

abstract backward(im)

Compute a backpropagation pass of the network. Sensitivity maps of each intermediate image are stored within the network.

Parameters

im – error gradient image (channels x rows x columns)

abstract gradient_zero()

Set all gradient variables to zero.

abstract gradient()

Compute gradient variables using computed sensitivity maps.

abstract getgradients()

Return a flat array with all gradient variables.

Returns

all gradient variables

updategradients(u)

Update variables of network within a thread.

Parameters

u – update variables

abstract updategradients_internal(u)

Update variables of network.

Parameters

u – update variables

abstract to_dict()

Return a dictionary containing all network variables and parameters.

Returns

all network variables and parameters

abstract load_dict(dct)

Set all network variables and parameters from dictionary.

Parameters

dct – all network variables and parameters

abstract classmethod from_dict(dct, gpu=True)

Initialize network and all network variables and parameters from dictionary.

Parameters

dct – all network variables and parameters

classmethod from_file(fn, gpu=True, groupname='network')

Initialize network and all network variables and parameters from file.

Parameters

• fn – filename

• gpu – (optional) whether to use GPU or CPU

to_file(fn, groupname='network')

Save all network variables and parameters to file.

Parameters

fn – filename

normalizeinout(datapoints)

Normalize input and output of network to zero mean and unit variance.

Parameters

datapoints – list of datapoints to compute normalization factors with.

class msdnet.network.MSDNet(d, dil, nin, nout, gpu=True)

Bases: msdnet.network.Network

Main implementation of a Mixed-Scale Dense network.

Parameters

• d – depth of network (width is always 1)

• dil – dilations.Dilations class defining dilations

• nin – number of input channels

• nout – number of output channels

• gpu – (optional) whether to use GPU or CPU

forward(im, returnoutput=True)

Compute a forward pass of the network.

Parameters

• im – input image (channels x rows x columns)

• returnoutput – whether to return the output image (default: True)

Returns

output image (channels x rows x columns)

backward(im, inputdelta=False)

Compute a backpropagation pass of the network. Sensitivity maps of each intermediate image are stored within the network.

Parameters

im – error gradient image (channels x rows x columns)

initialize()

Initialize network parameters.

gradient_zero()

Set all gradient variables to zero.

gradient()

Compute gradient variables using computed sensitivity maps.

filtergradient()

Compute filter gradient values.

getgradients()

Return a flat array with all gradient variables.

Returns

all gradient variables

updategradients_internal(u)

Update variables of network.

Parameters

u – update variables

setinputscale(gamma, offset)

Set input normalization values.

setoutputscale(gamma, offset)

Set output normalization values.

scaleinput()

Normalize input image.

scaleoutput()

Rescale output image.

scaleoutputback()

Rescale output image during backpropagation.

normalizeinput(datapoints)

Normalize input of network to zero mean and unit variance.

Parameters

datapoints – list of datapoints to compute normalization factors with.

normalizeoutput(datapoints)

Normalize output of network to zero mean and unit variance.

Parameters

datapoints – list of datapoints to compute normalization factors with.

to_dict()

Return a dictionary containing all network variables and parameters.

Returns

all network variables and parameters

load_dict(dct)

Set all network variables and parameters from dictionary.

Parameters

dct – all network variables and parameters

classmethod from_dict(dct, gpu=True)

Initialize network and all network variables and parameters from dictionary.

Parameters

dct – all network variables and parameters

class msdnet.network.SegmentationMSDNet(*args, **kwargs)

Bases: msdnet.network.MSDNet

Main implementation of a Mixed-Scale Dense network for segmentation.

Same parameters as MSDNet.

forward(im, returnoutput=True)

Compute a forward pass of the network.

Parameters

• im – input image (channels x rows x columns)

• returnoutput – whether to return the output image (default: True)

Returns

output image (channels x rows x columns)

normalizeoutput(datapoints)

Normalize output of network to zero mean and unit variance.

Parameters

datapoints – list of datapoints to compute normalization factors with.