GANEstimator with Combine Adversarial Loss

GANEstimator with Combine Adversarial Loss.py

def combined_loss(gan_model, **kwargs):
    """Wrapper function for combine adversarial loss, use as generator loss"""
    # Define non-adversarial loss - for example L1
    non_adversarial_loss = tf.losses.absolute_difference(
        gan_model.real_data, gan_model.generated_data)
    
    # Define generator loss
    generator_loss = tf.contrib.gan.losses.least_squares_generator_loss(
        gan_model,
        **kwargs)
    
    # The structure of kwargs changes between versions, better to add exception    
    try:
        add_summaries = kwargs['add_summaries']
    except:
        add_summaries = True
        
    # Combine these losses - you can specify more parameters
    # Exactly one of weight_factor and gradient_ratio must be non-None
    combined_loss = tf.contrib.gan.losses.wargs.combine_adversarial_loss(
        non_adversarial_loss,
        generator_loss,
        weight_factor=1.0,
        gradient_ratio=None,
        variables=gan_model.generator_variables,
        scalar_summaries=add_summaries,
        gradient_summaries=add_summaries)
    return combined_loss


gan_estimator = tf.contrib.gan.estimator.GANEstimator(
    model_dir,
    generator_fn=generator_fn,
    discriminator_fn=discriminator_fn,
    generator_loss_fn=combined_loss,
    discriminator_loss_fn=tf.contrib.gan.losses.least_squares_discriminator_loss,
    generator_optimizer=tf.train.AdamOptimizer(1e-4, 0.5),
    discriminator_optimizer=tf.train.AdamOptimizer(1e-4, 0.5))