niftynet.network.se_resnet module

class SE_ResNetDesc(bn, fc, conv1, blocks)

Bases: tuple

blocks

Alias for field number 3

bn

Alias for field number 0

conv1

Alias for field number 2

fc

Alias for field number 1

class SE_ResNet(num_classes, n_features=[16, 64, 128], n_blocks_per_resolution=1, w_initializer=None, w_regularizer=None, b_initializer=None, b_regularizer=None, acti_func='relu', name='SE_ResNet')

Bases: niftynet.network.base_net.BaseNet

### Description

implementation of Res-Net:

He et al., “Identity Mappings in Deep Residual Networks”, arXiv:1603.05027v3

### Building Blocks [CONV] - Convolutional layer, no activation, no batch norm (s)[DOWNRES] - Downsample residual block.

Each block is composed of a first bottleneck block with stride s, followed by n_blocks_per_resolution bottleneck blocks with stride 1. Bottleneck blocks include a squeeze-and-excitation block

[FC] - Fully connected layer with nr output channels == num_classes

### Diagram

INPUT –> [CONV] –>(s=1)[DOWNRES] –> (s=2)[DOWNRES] –> BN, ReLU, mean –> [FC] –> OUTPUT

### Constraints

__init__(num_classes, n_features=[16, 64, 128], n_blocks_per_resolution=1, w_initializer=None, w_regularizer=None, b_initializer=None, b_regularizer=None, acti_func='relu', name='SE_ResNet')

Parameters:

• num_classes – int, number of channels of output
• n_features – array, number of features per ResNet block
• n_blocks_per_resolution – int, number of BottleneckBlock per DownResBlock
• w_initializer – weight initialisation for network
• w_regularizer – weight regularisation for network
• b_initializer – bias initialisation for network
• b_regularizer – bias regularisation for network
• acti_func – ctivation function to use
• name – layer name

create()

Returns:tuple with batch norm layer, fully connected layer, first conv layer and all residual blocks

layer_op(images, is_training=True, **unused_kwargs)

Parameters:

• images – tensor, input to the network
• is_training – boolean, True if network is in training mode
• unused_kwargs – not in use

Returns:

tensor, output of the final fully connected layer

class BottleneckBlockDesc1(conv)

Bases: tuple

conv

Alias for field number 0

class BottleneckBlockDesc2(common_bn, conv, conv_shortcut)

Bases: tuple

common_bn

Alias for field number 0

conv

Alias for field number 1

conv_shortcut

Alias for field number 2

class BottleneckBlock(n_output_chns, stride, Conv, name='bottleneck')

Bases: niftynet.layer.base_layer.TrainableLayer

__init__(n_output_chns, stride, Conv, name='bottleneck')

Parameters:

• n_output_chns – int, number of output channels
• stride – int, stride to use in the convolutional layers
• Conv – layer, convolutional layer
• name – layer name

create(input_chns)

Parameters:input_chns – int, number of input channel Returns:tuple, with series of convolutional layers

layer_op(images, is_training=True)

Parameters:

• images – tensor, input to the BottleNeck block
• is_training – boolean, True if network is in training mode

Returns:

tensor, output of the BottleNeck block

class DownResBlockDesc(blocks)

Bases: tuple

blocks

Alias for field number 0

class DownResBlock(n_output_chns, count, stride, Conv, name='downres')

Bases: niftynet.layer.base_layer.TrainableLayer

__init__(n_output_chns, count, stride, Conv, name='downres')

Parameters:

• n_output_chns – int, number of output channels
• count – int, number of BottleneckBlocks to generate
• stride – int, stride for convolutional layer
• Conv – Layer, convolutional layer
• name – layer name

create()

Returns:tuple, containing all the Bottleneck blocks composing the DownRes block

layer_op(images, is_training)

Parameters:

• images – tensor, input to the DownRes block
• is_training – is_training: boolean, True if network is in training mode

Returns:

tensor, output of the DownRes block