Tests the down_shifted convolution for the vertical stack.

Tests the down_shifted convolution for the vertical stack.

Returns a boolean tensor representing (row_or_col % every == 0)

Tests the gated resnet layers for the horizontal stack.

Tests the gated resnet layers for the vertical stack.

Caches vertical stack hidden states. This avoids the need to pass the computed vertical stack in the feed_dict, which would involve CPU to GPU

Utility function to extract information about the current layer.

Creates all placeholders.

Pop off the oldest row of the cache to make space for the newest row of input.

utlity for keeping track of layer names

Creates a cache, which is used to avoid redundant computation.

Tests if horizontal stack outputs (one pixel at a time) of our code and OpenAI code are equal.

Computes cached convolutions, handling downsampling and upsampling.

Computes the conditional information for the resnet layer.

Creates the cache for the two deconv layers.

Creates and returns the variables used for a normal 2D convolution.

Creates and returns variables used for convolution.

Creates and returns the variables used for a 2D transposed convolution (deconvolution).

Sets up the conv tests by computing basic layer information.

Sets up both _test_*_equals() methods by initializing variables and outputs.

Tests if vertical stack outputs (one row at a time) of our code and OpenAI code are equal.

convolutional layer

transposed convolutional layer

Performs a convolution for the horizontal stack.

Performs a convolution for the vertical stack.

utility for retrieving polyak averaged params

utility for retrieving polyak averaged params

numerically stable log_sum_exp implementation that prevents overflow

Creates the model. Follows the same model_spec structure as the original PixelCNN++.

The down_shifted_conv2d adds a bias to the row of all zeros. This removes that bias.

Tests the sum of the vertical and horizontal stack.

fully connected layer

Tile images in a grid. If tile_shape is provided only as many images as specified in tile_shape will be included in the output.

numerically stable log_softmax implementation that prevents overflow

a network in network layer (1x1 CONV)

Scales all values in the ndarray ndar to be between 0 and 1

The down_shifted_conv2d adds a bias to the column of all zeros. This removes that bias.

Adam optimizer

like concatenated ReLU (http://arxiv.org/abs/1603.05201), but then with ELU

log-likelihood for mixture of discretized logistics, assumes the data has been rescaled to [-1,1] interval

Performs a transposed convolution for the horizontal stack.

Shift all rows down by one, using zeros as the first row and throwing away the last row.

Performs a transposed convolution for the vertical stack.

Performs gated resnet computations for the horizontal stack.

Performs gated resnet computations for the vertical stack.

Returns the matching conv or deconv function.

Returns an op to reset all created caches. Used between different generation calls.

Shift all columns right by one, using zeros as the first column and throwing away the last column.

Sums the vertical and horizontal stack.