Advanced Deep Learning with Keras

size=batch_size)
real_source = source_data[rand_indexes]
# generate a batch of fake target data fr real source data
fake_target = g_target.predict(real_source)
# combine real and fake into one batch
x = np.concatenate((real_target, fake_target))
# train the target discriminator using fake/real data
metrics = d_target.train_on_batch(x, valid_fake)
log = "%d: [d_target loss: %f]" % (step, metrics[0])
# generate a batch of fake source data fr real target data
fake_source = g_source.predict(real_target)
x = np.concatenate((real_source, fake_source))
# train the source discriminator using fake/real data
metrics = d_source.train_on_batch(x, valid_fake)
log = "%s [d_source loss: %f]" % (log, metrics[0])
Chapter 7
# train the adversarial network using forward and backward
# cycles. the generated fake source and target data attempts
# to trick the discriminators
x = [real_source, real_target]
y = [valid, valid, real_source, real_target]
metrics = adv.train_on_batch(x, y)
elapsed_time = datetime.datetime.now() - start_time
fmt = "%s [adv loss: %f] [time: %s]"
log = fmt % (log, metrics[0], elapsed_time)
print(log)
if (step + 1) % save_interval == 0:
if (step + 1) == train_steps:
show = True
else:
show = False
test_generator((g_source, g_target),
(test_source_data, test_target_data),
step=step+1,
titles=titles,
dirs=dirs,
show=show)
# save the models after training the generators
g_source.save(model_name + "-g_source.h5")
g_target.save(model_name + "-g_target.h5")
[ 223 ]