Deep Learning for Computer Vision

InfoGAN

InfoGAN can generate images of the required label without any explicit supervised training. The infogan_model takes unstructured and structured input, as shown in the following code:

info_gan = tf.contrib.gan.infogan_model(    get_generator,    get_discriminator,    real_images,    unstructured_input,    structured_input)loss = tf.contrib.gan.gan_loss(    info_gan,    gradient_penalty_weight=1,    gradient_penalty_epsilon=1e-10,    mutual_information_penalty_weight=1)

The loss is defined with a penalty as the training is unstable. Adding the penalty provides more stability during the training.