Bottom-up deterministic layer for inference, roughly the same as the small deterministic Resnet in top-down layers. Consists of a sequence of

Wrapper for DataLoaders. Data attributes: - train: DataLoader object for training set - test: DataLoader object for test set - data_s

Assume input data to be originally uint8 (0, ..., 255) and then rescaled by 1/255: discrete values in {0, 1/255, ..., 255/255}. If using

Sampling and loss computation are based on the original tf code. Assume input data to be originally uint8 (0, ..., 255) and then rescaled

Double the number of channels through a convolutional layer, then use half the channels as gate for the other half.

Experiment manager. Data attributes: - 'args': argparse.Namespace containing all config parameters. When initializing this object,

Merge two 4D input tensors by concatenating along dim=1 and passing the result through 1) a convolutional 1x1 layer, or 2) a residual block

Transform input parameters to q(z) with a convolution, optionally do the same for p(z), then sample z ~ q(z) and return conv(z). If q's

Residual block with 2 convolutional layers. Input, intermediate, and output channels are the same. Padding is always 'same'. The 2 convol

By default for now simply a merge layer.

Top-down layer, including stochastic sampling, KL computation, and small deterministic ResNet with upsampling. The architecture when doi

Residual block that takes care of resampling steps (each by a factor of 2). The mode can be top-down or bottom-up, and the block does up- an