Neural Nets

Fully-connected nets

class pyroved.nets.fcEncoderNet(in_dim, latent_dim=2, num_classes=0, hidden_dim=128, num_layers=2, activation='tanh', softplus_out=True, flat=True)

Bases: torch.nn.Module

Standard fully-connected encoder NN for VAE. The encoder outputs mean and standard evidation of the encoded distribution.

class pyroved.nets.jfcEncoderNet(in_dim, latent_dim=2, discrete_dim=0, hidden_dim=128, num_layers=2, activation='tanh', softplus_out=True, flat=True)

Bases: torch.nn.Module

Fully-connected encoder for joint VAE. The encoder outputs mean, standard evidation and class probabilities.

class pyroved.nets.fcDecoderNet(out_dim, latent_dim, num_classes=0, hidden_dim=128, num_layers=2, activation='tanh', sigmoid_out=True, unflat=True)

Bases: torch.nn.Module

Standard fully-connected decoder for VAE

class pyroved.nets.sDecoderNet(out_dim, latent_dim, num_classes=0, hidden_dim=128, num_layers=2, activation='tanh', sigmoid_out=True, unflat=True)

Bases: torch.nn.Module

Spatial generator (decoder) network with fully-connected layers

Convolutional nets

class pyroved.nets.convEncoderNet(input_dim, input_channels=1, latent_dim=2, layers_per_block=None, hidden_dim=32, batchnorm=True, activation='lrelu', softplus_out=True, pool=True)

Bases: torch.nn.Module

Standard convolutional encoder

class pyroved.nets.convDecoderNet(latent_dim, output_dim, output_channels=1, layers_per_block=None, hidden_dim=96, batchnorm=True, activation='lrelu', sigmoid_out=True, upsampling_mode='bilinear')

Bases: torch.nn.Module

Standard convolutional decoder

class pyroved.nets.FeatureExtractor(ndim, input_channels=1, layers_per_block=None, nfilters=32, batchnorm=True, activation='lrelu', pool=True)

Bases: torch.nn.Sequential

Convolutional feature extractor

class pyroved.nets.Upsampler(ndim, input_channels=96, layers_per_block=None, output_channels=1, batchnorm=True, activation='lrelu', upsampling_mode='bilinear')

Bases: torch.nn.Sequential

Convolutional upsampler

class pyroved.nets.ConvBlock(ndim, nlayers, input_channels, output_channels, kernel_size=3, stride=1, padding=1, batchnorm=False, activation='lrelu', pool=False)

Bases: torch.nn.Module

Creates a block of layers each consisting of convolution operation, (optional) nonlinear activation and (optional) batch normalization

class pyroved.nets.UpsampleBlock(ndim, input_channels, output_channels, scale_factor=2, mode='bilinear')

Bases: torch.nn.Module

Upsampling performed using bilinear or nearest-neigbor interpolation followed by 1-by-1 convolution, which an be used to reduce a number of feature channels