Pytorch models

class ConvBlock(layout, in_channels=None, filters=None, kernel_size=3, stride=1, padding='same', rate=0.1, activation=None)

Convolutional block.

Parameters

• in_channels (int) – Input channels size.

• layout (str) – Layout of layers. Can contain “c” for convolution, “a” for activation, “n” for batchnorm, “p” for maxpooling, “d” for dropout. E.g. of layout: “ccnapd”.

• filters (int, list or None) – Number of filters for convolutions. Can be a single number (all convolutions will have the same number of filters), a list of the same length as a count of letters “c” in the layout, or None if the layout contains no “c”.

• rate (float) – Dropout rate parameter. Default to 0.1.

• activation (function) – Activation function. If not specified activation is tf.nn.elu.

forward(x)

Forward pass.

training = None

class Unet(in_channels, depth, init_filters, kernel_size=3, output=None, norm=True)

U-net implementation.

Parameters

• in_channels (int) – Input channels size.

• depth (int) – Depth of the U-net.

• init_filters (int) – Number of filters in the first conv block.

• output (dict) – Output conv block config.

forward(x)

Forward pass.

training = None

class VAE(in_channels, filters_enc, filters_dec, z_dim, output, kernel_size=3, norm=True, variational=True)

Convolutional VAE model.

Parameters

• in_channels (int) – Input channels size.

• filters_enc (array) – Sequence of filters for encoder.

• filters_enc – Sequence of filters for decoder.

• z_dim (int) – Number of filters in the latent space.

• output (dict) – Output conv block config for decoder.

• kernel_size (int) – Kernel size in convolution layers. Default 3.

• norm (bool) – Include normalization layers. Default True.

• variational (bool) – If False, implemnt ordinary AE scheme. Default True.

decode(z)

Decoder.

encode(x)

Encoder.

forward(x)

Forward pass.

reparameterize(mu, logvar)

Reparametrization trick.

training = None