Applies a gated activation to the convolutional volume. Note that this activation divides the number of channels by 2. Parameters -------

Trains model on the data given in data_loader. Parameters ---------- data_loader : torch.utils.data.DataLoader instance

Returns conditional pixels obtained from masking the data in batch with mask and appending the mask. E.g. if the input has size (N, C, H, W) t

Returns a tensor of shape (batch_size, 1, img_size[1], img_size[2]) containing masks which were generated according to mask_type and m

Generates samples from a GatedPixelCNN or GatedPixelCNNRGB. Parameters ---------- device : torch.device instance num

Returns a batch of random masks. Parameters ---------- img_size : see single_random_mask num_visibles : int or list of ints

Returns random mask where 0 corresponds to a hidden value and 1 to a visible value. Shape of mask is same as img_size. Parameters -------

CelebA dataloader with square images. Note original CelebA images have shape (218, 178), this dataloader center crops these images to be (178, 178

Forward method is implemented in child class (GatedPixelCNN or GatedPixelCNNRGB).

x : torch.Tensor Image to predict logits for. x_cond : torch.Tensor Image containing pixels to be conditione

Generates image completions based on the images in batch masked by the masks in mask_descriptors. This will generate batch.size(0) * len(mask_

Returns a quantization function which can be used to set the number of colors in an image. Parameters ---------- num_colors : int

Returns repeated conditional pixels. Parameters ---------- batch : torch.Tensor Shape (1, num_channels, width, height) mask

Helper function that initializes an appropriate model based on the input arguments. Parameters ---------- img_size : tuple of ints

MNIST dataloader with (28, 28) images. Parameters ---------- batch_size : int num_colors : int Number of colors to quantize

Masks all the output except the |num_rows| lowest rows (in the height dimension). Parameters ---------- img_size : see single_random_

Masks all the output except the num_pixels by num_pixels central square of the image. Parameters ---------- img_size : see single_ran

Masks all the output except the num_pixels thick edge of the image. Parameters ---------- img_size : see single_random_mask num_pixe

Masks all the pixels except the left side of the image. Parameters ---------- img_size : see single_random_mask num_cols : int

Generates batch of masks with multiple random connected blobs.

Returns a batch of masks with random rectangles of visible pixels. Parameters ---------- img_size : see single_random_mask max_heigh

Masks all the pixels except the right side of the image. Parameters ---------- img_size : see single_random_mask num_cols : int

Masks all the output except the |num_rows| highest rows (in the height dimension). Parameters ---------- img_size : see single_random

Returns model, data_loader and mask_descriptor of trained model. Parameters ---------- directory : string Directory where ex

Calculates log likelihood of samples under model. Parameters ---------- device : torch.device instance samples : tor

Calculates log likelihood of samples under model. Parameters ---------- samples : torch.Tensor Batch of images. S

Generates masks with multiple random connected blobs. Parameters ---------- max_num_blobs : int Maximum number of blobs. Number o

Creates a plot of pixel probabilities with the conditional pixels from the original image overlayed. Note this function only works for binary

Generates masks with random connected blobs. Parameters ---------- img_size : see single_random_mask num_iter : int Number o

Returns a mask with a random rectangle of visible pixels. Parameters ---------- img_size : see single_random_mask max_height : int

Generate conditional samples from the model. The number of samples generated will be equal to the batch size of x_cond. Parameters

Sorts samples by their log likelihoods and creates an image representing the log likelihood of each sample as a box with color and size proportion

Gated PixelCNN convolutional block. Note the number of input and output channels must be the same, unless restricted is True. Paramet

Gated PixelCNN convolutional block for RGB images. Note the number of input and output channels must be the same, unless restricted is True.

Takes as input a float tensor with values in the 0 - 1 range and outputs a long tensor with integer values corresponding to each quant

Samples from prior model without conditioning.

Parameters ---------- data_loader : torch.utils.data.DataLoader instance epochs : int Number of epochs t