I am working with the improved version of Wasserstein GANs, for generated images when given as input some features X_{S} coupled together with noise vector z. Furthermore, I would like to generate samples from specific classes. Therefore, a classifier was added to the standard training of the whole system. My code for building the whole system is the following:
class WGANGP():
def __init__(self):
self.target_mod = ".."
self.learning_param = 0.0001
self.no_input_feats = ...
# Following parameter and optimizer set as recommended in paper
self.n_critic = 15
optimizer = RMSprop(lr=self.learning_param)
self.img_rows = ...
self.img_cols = ...
self.channels = 3
self.img_shape = (self.img_rows, self.img_cols, self.channels)
self.latent_dim = ...
# Build the generator and critic
self.generator = build_generator(self.latent_dim, self.channels)
#self.generator = self.build_generator_old()
self.critic = build_critic(self.img_shape)
#-------------------------------
# Construct Computational Graph
# for the Critic
#-------------------------------
# Freeze generator's layers while training critic
self.generator.trainable = False
# Image input (real sample)
real_img = Input(shape=self.img_shape)
# Noise input
z_disc = Input(shape=(self.latent_dim,))
# Generate image based of noise (fake sample)
fake_img = self.generator(z_disc)
# Discriminator determines validity of the real and fake images
fake, aux1 = self.critic(fake_img)
valid, aux2 = self.critic(real_img)
# Construct weighted average between real and fake images
interpolated_img = RandomWeightedAverage()([real_img, fake_img])
# Determine validity of weighted sample
validity_interpolated, aux3 = self.critic(interpolated_img)
# Use Python partial to provide loss function with additional
# 'averaged_samples' argument
partial_gp_loss = partial(self.gradient_penalty_loss,
averaged_samples=interpolated_img)
partial_gp_loss.__name__ = 'gradient_penalty'
# Keras requires function names
self.critic_model = Model(inputs=[real_img, z_disc],
outputs=[valid, fake, validity_interpolated, aux1])
self.critic_model.compile(loss=[self.wasserstein_loss,
self.wasserstein_loss,
partial_gp_loss,
'categorical_crossentropy'],
optimizer=optimizer,
metrics=['accuracy'],
loss_weights=[1, 1, 5, 1])
#-------------------------------
# Construct Computational Graph
# for Generator
#-------------------------------
# For the generator we freeze the critic's layers
self.critic.trainable = False
self.generator.trainable = True
# Sampled noise for input to generator
z_gen = Input(shape=(self.latent_dim,))
# Generate images based of noise
img = self.generator(z_gen)
# Discriminator determines validity
valid = self.critic(img)
# Defines generator model
self.generator_model = Model(z_gen, valid)
self.generator_model.compile(loss=self.wasserstein_loss, optimizer=optimizer)
When I am printing the critic_model.metric_names I am receiving the following:
['loss', 'model_2_loss', 'model_2_loss', 'model_2_loss', 'model_2_loss', 'model_2_acc', 'model_2_acc_1', 'model_2_acc_2', 'model_2_acc_3']
Can anyone help me understand what these names stand for?
The answer is right here:
self.critic_model = Model(inputs=[real_img, z_disc],
outputs=[valid, fake, validity_interpolated, aux1]) #<- 4 outputs
#4 model losses + 1 total loss:
self.critic_model.compile(loss=[self.wasserstein_loss, #loss for output 0
self.wasserstein_loss, #loss for output 1
partial_gp_loss, #loss for output 2
'categorical_crossentropy'] #loss for output 3
optimizer=optimizer,
metrics=['accuracy'], #replicated, one for each output
loss_weights=[1, 1, 5, 1])
Your model clearly has 4 outputs, and you defined one loss per output. Whenever you have multiple losses, Keras will sum a total loss for you, so:
'loss' is the total loss (sum of all losses for this model) The other 4 'model_2_loss' are, in order:
self.wasserstein_loss, for the first output valid self.wasserstein_loss, for the second output fake partial_gp_loss, for validity_interpolated 'categorical_crossentropy' for aux1 For the metrics, since you defined only one, the system replicated this same metric for each of the model's outputs:
'model_2_acc', metric for valid 'model_2_acc_1', metric for fake 'model_2_acc_2', metric for validity_interpolated 'model_2_acc_3', metric for aux1 For better loss names, you should add names to the outputs of the models, to the losses, etc., wherever it's possible to add a name parameter.
Some operations accept names, such as: How can I set the name of my loss operation in Tensorflow?
Losses in newer versions, created as objects, also accept names: https://www.tensorflow.org/api_docs/python/tf/keras/losses/CategoricalCrossentropy
Models accept names, such as:
self.critic_model = Model(inputs=[real_img, z_disc],
outputs=[valid, fake, validity_interpolated, aux1],
name='critic_model')
Layers accept names, so you should have a name for the output layer of each model to keep track of things better.
I'm not totally familiar with the new eager mode execution concept, but you might try to add a name also whenever you call a model... not sure whether this is possible.