I am running to this error while executing gradient penalty for WGAN model
File c:\Users\Mega PC\Desktop\Internship\GANs\GANs_for_Credit_Card_Data-master\GAN_171103.py:405, in adversarial_training_WGAN(arguments, train, data_cols, label_cols, seed, starting_step)
403 # update f by taking an SGD step on mini-batch loss LD(f)
404 disc_optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate, beta_1=0.5, beta_2=0.9)
→ 405 disc_optimizer.minimize(_disc_loss, var_list=discriminator_model.trainable_weights)
407 sess = K.get_session()
409 # Compile models
File c:\Users\Mega PC\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\optimizers\optimizer.py:542, in _BaseOptimizer.minimize(self, loss, var_list, tape)
521 def minimize(self, loss, var_list, tape=None):
522 “”“Minimize loss by updating var_list.
523
524 This method simply computes gradient using tf.GradientTape and calls
(…)
540 None
541 “””
→ 542 grads_and_vars = self.compute_gradients(loss, var_list, tape)
543 self.apply_gradients(grads_and_vars)
File c:\Users\Mega PC\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\optimizers\optimizer.py:261, in _BaseOptimizer.compute_gradients(self, loss, var_list, tape)
243 “”“Compute gradients of loss on trainable variables.
244
245 Args:
(…)
258 gradient can be None.
259 “””
260 if not callable(loss) and tape is None:
→ 261 raise ValueError(
262 “tape is required when a Tensor loss is passed. "
263 f"Received: loss={loss}, tape={tape}.”
264 )
265 if tape is None:
266 tape = tf.GradientTape()
ValueError: tape is required when a Tensor loss is passed. Received: loss=Tensor(“add_31:0”, shape=(), dtype=float32), tape=None.
def adversarial_training_WGAN(arguments, train, data_cols, label_cols=[], seed=0, starting_step=0):
[rand_dim, nb_steps, batch_size, k_d, k_g, critic_pre_train_steps, log_interval, learning_rate, base_n_count,
data_dir, generator_model_path, discriminator_model_path, loss_pickle_path, show] = arguments
np.random.seed(seed) # set random seed
data_dim = len(data_cols)
print('data_dim: ', data_dim)
print('data_cols: ', data_cols)
label_dim = 0
with_class = False
if len(label_cols) > 0:
with_class = True
label_dim = len(label_cols)
print('label_dim: ', label_dim)
print('label_cols: ', label_cols)
# Define network models
tf.random.set_seed(seed) # set random seed for TensorFlow 2.x
K.set_learning_phase(1) # 1 = train
if with_class:
cache_prefix = 'WCGAN'
generator_model, discriminator_model, combined_model = define_models_CGAN(rand_dim, data_dim, label_dim,
base_n_count, type='Wasserstein')
else:
cache_prefix = 'WGAN'
generator_model, discriminator_model, combined_model = define_models_GAN(rand_dim, data_dim, base_n_count,
type='Wasserstein')
# Construct computation graph for calculating the gradient penalty (improved WGAN) and training the discriminator
_z = tf.keras.Input(shape=(rand_dim,))
_labels = None
if with_class:
_x = tf.keras.Input(shape=(data_dim + label_dim))
_labels = tf.keras.Input(shape=(label_dim,))
_g_z = generator_model(inputs=[_z, _labels])
else:
_x = tf.keras.Input(shape=(data_dim,))
_g_z = generator_model(_z)
epsilon = tf.keras.Input(shape=(batch_size, 1))
x_hat = epsilon * _x + (1.0 - epsilon) * _g_z
x_hat_reshaped = tf.reshape(x_hat, shape=(-1, 4))
gradients = tf.gradients(discriminator_model(x_hat_reshaped), [x_hat_reshaped])
# gradients = tf.gradients(discriminator_model(x_hat), [x_hat])
_gradient_penalty = 10.0 * tf.square(tf.norm(gradients[0], ord=2) - 1.0)
# Calculate discriminator's loss
_disc_loss_generated = em_loss(tf.ones(batch_size), discriminator_model(_g_z))
_disc_loss_real = em_loss(tf.ones(batch_size), discriminator_model(_x))
_disc_loss = _disc_loss_generated - _disc_loss_real + _gradient_penalty
# update f by taking an SGD step on mini-batch loss LD(f)
disc_optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate, beta_1=0.5, beta_2=0.9)
disc_optimizer.minimize(_disc_loss, var_list=discriminator_model.trainable_weights)
sess = K.get_session()
# Compile models
adam = tf.keras.optimizers.Adam(lr=learning_rate, beta_1=0.5, beta_2=0.9)
discriminator_model.trainable = False
combined_model.compile(optimizer=adam, loss=[em_loss])
combined_loss, disc_loss_generated, disc_loss_real, xgb_losses = [], [], [], []
model_components = [cache_prefix, with_class, starting_step,
train, data_cols, data_dim,
label_cols, label_dim,
generator_model, discriminator_model, combined_model,
rand_dim, nb_steps, batch_size,
k_d, k_g, critic_pre_train_steps, log_interval, learning_rate, base_n_count,
data_dir, generator_model_path, discriminator_model_path,
_z, _x, _labels, _g_z, epsilon, x_hat_reshaped, gradients, _gradient_penalty,
_disc_loss_generated, _disc_loss_real, _disc_loss, disc_optimizer,
show,
combined_loss, disc_loss_generated, disc_loss_real, xgb_losses]
if show:
print(generator_model.summary())
print(discriminator_model.summary())
print(combined_model.summary())
if loss_pickle_path:
print('Loading loss pickles')
[combined_loss, disc_loss_generated, disc_loss_real, xgb_losses] = pickle.load(open(loss_pickle_path, 'rb'))
if generator_model_path:
print('Loading generator model')
generator_model.load_weights(generator_model_path)
if discriminator_model_path:
print('Loading discriminator model')
discriminator_model.load_weights(discriminator_model_path)
else:
print('Pre-training the critic...')
K.set_learning_phase(1) # 1 = train
for i in range(critic_pre_train_steps):
if i % 20 == 0:
print('Step: {} of {} critic pre-training.'.format(i, critic_pre_train_steps))
loss = train_discriminator_step(model_components, seed=i)
print('Last batch of critic pre-training disc_loss: {}.'.format(loss))
model_components = [cache_prefix, with_class, starting_step,
train, data_cols, data_dim,
label_cols, label_dim,
generator_model, discriminator_model, combined_model,
rand_dim, nb_steps, batch_size,
k_d, k_g, critic_pre_train_steps, log_interval, learning_rate, base_n_count,
data_dir, generator_model_path, discriminator_model_path,
_z, _x, _labels, _g_z, epsilon, x_hat_reshaped, gradients, _gradient_penalty,
_disc_loss_generated, _disc_loss_real, _disc_loss, disc_optimizer,
show,
combined_loss, disc_loss_generated, disc_loss_real, xgb_losses]
[combined_loss, disc_loss_generated, disc_loss_real, xgb_losses] = training_steps_WGAN(model_components)