def __init__(self,
generator,
discriminator,
content_loss='VGG54',
learning_rate=PiecewiseConstantDecay(boundaries=[100000],
values=[1e-4, 1e-5])):
if content_loss == 'VGG22':
self.vgg = srgan.vgg_22()
elif content_loss == 'VGG54':
self.vgg = srgan.vgg_54()
else:
raise ValueError("content_loss must be either 'VGG22' or 'VGG54'")
self.content_loss = content_loss
self.generator = generator
self.discriminator = discriminator
self.generator_optimizer = Adam(learning_rate=learning_rate)
self.discriminator_optimizer = Adam(learning_rate=learning_rate)
self.binary_cross_entropy = BinaryCrossentropy(from_logits=False)
self.mean_squared_error = MeanSquaredError()
log_param("Model", "Pre-trained SRGAN")
log_param("Learning rate", learning_rate)
log_param("Content loss", content_loss)
def content_loss(hr, sr):
vgg = srgan.vgg_54()
sr = preprocess_input(sr)
hr = preprocess_input(hr)
sr_features = vgg(sr)
hr_features = vgg(hr)
return mean_squared_error(hr_features, sr_features)
def __init__(self,
generator,
discriminator,
content_loss='VGG54',
learning_rate=PiecewiseConstantDecay(boundaries=[100000],
values=[1e-4, 1e-5])):
if content_loss == 'VGG22':
self.vgg = srgan.vgg_22()
elif content_loss == 'VGG54':
self.vgg = srgan.vgg_54()
else:
raise ValueError("content_loss must be either 'VGG22' or 'VGG54'")
self.content_loss = content_loss
self.teacher = generator
self.teacher.load_model_from_file('weights/srgan/gan_generator.h5')
self.student = generator(num_res_blocks=6)
self.discriminator = discriminator
self.generator_optimizer = Adam(learning_rate=learning_rate)
self.discriminator_optimizer = Adam(learning_rate=learning_rate)
self.binary_cross_entropy = BinaryCrossentropy(from_logits=False)
self.mean_squared_error = MeanSquaredError()