strided 2-D convolution with 'REFLECT' padding. Args: inputs: A 4-D tensor of size [batch, height, width, channel] num_outputs: A

deconvolution alternatively with the conv2d_transpose, where we first resize the inputs, and then convolve the results, see http://distill

helper fn that provides the full image filenames from the dataset_dir

help fn that provides numpy image coding utilities

Configures the learning rate. Args: num_samples_per_epoch: The number of samples in each epoch of training global_step: The globa

Configures the optimizer used for training. Args: learning_rate: A scalar or 'Tensor' learning rate Returns: An instance of

Returns a function to warm-start the training. Note that the init_fn is only run when initializing the models during the very first global st

Returns a list of variables to train. Args: A list of variables to train by the optimizer.

build the training graph for the training of the hierarchical autoencoder

compute the content features from the end_point dict

compute the gram matrix for a layer of feature the gram matrix is normalized with respect to the samples and the dimensions of the input

compute the style features from the end_point dict

compute the gram matrix for a layer of feature

compute the dict of layer-wise image features from a given list of networks Args: inputs: the inputs image should be normalized between [-1

hierarchical autoencoder for content reconstruction

works for one single image with dynamic shapes

style decorator for high-level feature interaction Args: content_features: a tensor of size [batch_size, height, width, channel]

transfer the content image by style images Args: inputs: input images [batch_size, height, width, channel] styles: a

adaptively transform the content feature by inverse instance normalization based on the 2nd order statistics of the style feature

compute the approximate style loss Args: style_features: a dict of the sufficient statistics of the feature maps of the style

compute the content and style features from normalized image Args: inputs: input tensor of size [batch, height, width, channel] netwo

compute the content loss Args: content_features: a dict of the features of the input image target_features: a dict of the features of

compute the style loss Args: style_features: a dict of the Gram matrices of the style image target_features: a dict of the Gram m

compute the total variation loss L1 norm

compute the total variation loss

rescale the images so that their magnitude ranging from [-1, 1]

customized leaky ReLU activation function https://github.com/tensorflow/tensorflow/issues/4079

works for one single image with dynamic shapes

Residual block version 2 for downsampling, with preactivation Args: inputs: a tensor of size [batch, height, width, channel] dept

Residual block version 2 for upsampling, with preactivation Args: inputs: a tensor of size [batch, height, width, channel] depth:

Residual block version 0, the input and output has the same depth Args: inputs: a tensor of size [batch, height, width, channel] dept

Oxford Net VGG 16-Layers version D Example. Note: All the fully_connected layers have been transformed to conv2d layers. To use in classi

Oxford Net VGG 19-Layers version E Example. Note: All the fully_connected layers have been transformed to conv2d layers. To use in classi

Oxford Net VGG 11-Layers version A Example. Note: All the fully_connected layers have been transformed to conv2d layers. To use in classi

Defines the VGG arg scope. Args: weight_decay: The l2 regularization coefficient. Returns: An arg_scope.

construct the decoder network with additional feature combination Args: inputs: input features [batch_size, height, width, channel]

construct the decoder network for the vgg models Args: inputs: input features [batch_size, height, width, channel] network_name:

construct the decoder network with additional feature combination Args: inputs: input features [batch_size, height, width, channel]

transform the content feature based on the whitening and colorization transform