def train_gan(wgan,
batch_gen,
noise_shapes,
steps_per_epoch,
num_epochs,
application="mch"):
img_shape = batch_gen.img_shape
noise_gen = noise.NoiseGenerator(noise_shapes(img_shape),
batch_size=batch_gen.batch_size)
for epoch in range(num_epochs):
print("Epoch {}/{}".format(epoch + 1, num_epochs))
wgan.train(batch_gen, noise_gen, steps_per_epoch, training_ratio=5)
plots.plot_samples(
wgan.gen,
batch_gen,
noise_gen,
application=application,
out_fn="../figures/progress_{}.pdf".format(application))
return wgan
def train_gan(gan, batch_gen, noise_gen, num_epochs=1, steps_per_epoch=1, plot_fn=None):
gan.fit_generator(batch_gen, noise_gen, num_epochs=num_epochs,
steps_per_epoch=steps_per_epoch)
plots.plot_samples(gan.gen, batch_gen, noise_gen, out_fn=plot_fn)
fs_loss = f_gan.train_generator() # f stacked
fd_loss, fd_acc = f_gan.train_discriminator() # f discriminator
gs_loss = g_gan.train_generator() # g stacked
gd_loss, gd_acc = g_gan.train_discriminator() # g discriminator
leftinv_loss = cycle_gan.train_left_inverse()
rightinv_loss = cycle_gan.train_right_inverse()
if train.log:
print(f'{fs_loss:1.5f} {gs_loss:1.5f} {fd_acc:.5f} {gd_acc:.5f}')
if train.make_samples:
g_sample = f_gan.generate_sample(5)
train.append_sample(g_sample,
'samples/sample_{gen:03.0f}_{i:02.0f}.png')
train.append_stats(fs_loss, fd_loss, fd_acc, gs_loss, gd_loss, gd_acc,
leftinv_loss, rightinv_loss)
except KeyboardInterrupt:
pass
else:
train.compile_records()
plot_samples(train.samples)
plot_stats(train.stats)
save_model(f_gan.generator)
save_model(g_gan.generator)