calphagan

Brief Introduction

We run this script under TensorFlow 1.4 and the TensorLayer 1.8.0+. The proposed model integrates conditional GAN and recently introduced alpha-GAN architecture, aiming at solving the image generation problem conditioned on brain signal.

alpha-GAN: Variational Approaches for Auto-Encoding Generative Adversarial Networks

conditional GAN: Conditional Generative Adversarial Nets

Network Architecture

Encoder

CodeDiscriminator

Generator

Discriminator

Data

Save image data into TFRecord format

In this script, we read the image data, and save them into TFRecord format using the following method.

load_and_save_to_tfrecord(data_dir, save_dir, name)

data_dir should contain the folders named by the class labels, and each folder contains the images in JPEG format. Finally the data would be saved into the save_dir with the given name.

Read TFRecord during training

The data is read from TFRecord from following API.

input_batch(filename, batch_size, num_epochs, shuffle_size, is_augment)

filename is the TFRecord file storing all the images and corresponding label. is_augment is a boolean type parameter to choose use augmentation method or not. The return of this method is a batch of images and their labels.

Run

To successfully run the code, please firstly set up all the configurations in god_config.py, such as the TFRecord file directory

config.data_tfrecord_dir = './train/cifar10_labeled.tf'

and other tunable hyper-parameters.

There are three modes to train the networks and evaluate the results respectively.

Train

The train mode is used to train the whole network, including generator, encoder, discriminator and code discriminator. Simply using the following command on terminal.

python main.py

Encode

The encode mode uses well-trained encoder to encode the provided images under train_samples.

images_dir = './train_samples'

To run this mode, you can simply use the following command on terminal.

python main.py --mode=encode

Generate

The generate mode uses well-trained generator to generate the reconstructed images from given predicted features and labels, which is stored in following path

'./results/encoded_feat.pkl'

To run this mode, you can simply use the following command on terminal.

python main.py --mode=gen

Results

Test on CIFAR-10

Test on random constructed dataset