ActNorm ([axis, value_ndims, initialized, …]) |
ActNorm proposed by (Kingma & Dhariwal, 2018). |
BaseFlow (x_value_ndims[, y_value_ndims, …]) |
The basic class for normalizing flows. |
BaseLayer ([name, scope]) |
Base class for all neural network layers. |
CouplingLayer (shift_and_scale_fn[, axis, …]) |
A general implementation of the coupling layer (Dinh et al., 2016). |
FeatureMappingFlow (axis, value_ndims, **kwargs) |
Base class for flows mapping input features to output features. |
FeatureShufflingFlow ([axis, value_ndims, …]) |
An invertible flow which shuffles the order of input features. |
InvertFlow (flow[, name, scope]) |
Turn a BaseFlow into its inverted flow. |
InvertibleActivation |
Base class for intertible activation functions. |
InvertibleActivationFlow (activation, value_ndims) |
A flow that converts a InvertibleActivation into a flow. |
InvertibleConv2d ([channels_last, …]) |
Invertible 1x1 2D convolution proposed in (Kingma & Dhariwal, 2018). |
InvertibleDense ([strict_invertible, …]) |
Invertible dense layer, modified from the invertible 1x1 2d convolution proposed in (Kingma & Dhariwal, 2018). |
LeakyReLU ([alpha]) |
Leaky ReLU activation function. |
MultiLayerFlow (n_layers, **kwargs) |
Base class for multi-layer normalizing flows. |
PixelCNN2DOutput (vertical, horizontal) |
The output of a PixelCNN 2D layer, including tensors from the vertical and horizontal convolution stacks. |
PlanarNormalizingFlow ([w_initializer, …]) |
A single layer Planar Normalizing Flow (Danilo 2016) with tanh activation function, as well as the invertible trick. |
ReshapeFlow (x_value_ndims, y_value_shape[, …]) |
A flow which reshapes the last x_value_ndims of x into y_value_shape. |
SequentialFlow (flows[, name, scope]) |
Compose a large flow from a sequential of BaseFlow . |
SpaceToDepthFlow (block_size[, …]) |
A flow which computes y = space_to_depth(x) , and conversely x = depth_to_space(y) . |
SplitFlow (split_axis, left[, join_axis, …]) |
A flow which splits input x into halves, apply different flows on each half, then concat the output together. |