def plotting_process(vae, decoder, vaegan, input_test, run_folder, latent_dim, w_filename):
# Load weights
vaegan.load_weights(w_filename)
# random latent space samples to plot
noise = np.random.normal(0, 1, (5 * 5, latent_dim))
sample_images(decoder, os.path.join(run_folder, "sample-model"), noise, True)
# real vs reconstruction plot
plotRealReconstruction(vae, input_test, 1, 5, os.path.join(run_folder, "realReconstructedSamples"), True)
def run_training(train_dataset, encoder, decoder, discriminator, vaegan, optimizers, gamma, noise, run_folder, label_smooth_r, label_smooth_f):
enc_loss_hist = np.array([])
dec_loss_hist = np.array([])
dis_loss_hist = np.array([])
for epoch in range(NO_EPOCHS):
l = train(train_dataset, encoder, decoder, discriminator, optimizers, gamma, label_smooth_r, label_smooth_f)
if (epoch+1) % 1 == 0:
print('epoch:', epoch+1, 'kl_loss:', l[0], 'rec_loss:', l[1], 'gen_loss:', l[2], 'dis_loss:', l[3])
enc_loss_hist = np.append(enc_loss_hist, (l[0]+l[1]))
dec_loss_hist = np.append(dec_loss_hist, (l[1]-l[3]))
dis_loss_hist = np.append(dis_loss_hist, l[3])
sample_images(decoder, os.path.join(run_folder,"epochSamples","sample-%d" % (epoch+1)), noise, True)
vaegan.save_weights(os.path.join(run_folder, "weights", "weights-%d.h5" % (epoch+1)))
return enc_loss_hist, dec_loss_hist, dis_loss_hist
def train(gan_model, gen_model, dis_model, x_train, batch_size, no_epochs,
run_folder, noise, z_dim, label_noise_pct, label_smooth_r,
label_smooth_f):
"""
train the gan model
"""
gen_loss_hist = np.array([])
dis_loss_hist = np.array([])
dis_loss_real_hist = np.array([])
dis_loss_fake_hist = np.array([])
nb_batch = x_train.shape[0] // batch_size
idx_train = np.arange(len(x_train))
for epoch in range(1, no_epochs + 1):
np.random.shuffle(idx_train)
genera_losses = np.array([])
discri_losses = np.array([])
discri_loss_r = np.array([])
discri_loss_f = np.array([])
for i in range(nb_batch):
batch_x_train = x_train[idx_train[i * batch_size:(i + 1) *
batch_size]]
d = train_discriminator(gen_model, dis_model, batch_x_train,
len(batch_x_train), z_dim, label_noise_pct,
label_smooth_r, label_smooth_f)
g = train_generator(gan_model, batch_size, z_dim)
genera_losses = np.append(genera_losses, g)
discri_losses = np.append(discri_losses, d[0])
discri_loss_r = np.append(discri_loss_r, d[1])
discri_loss_f = np.append(discri_loss_f, d[2])
gen_loss_hist = np.append(gen_loss_hist, np.mean(genera_losses))
dis_loss_hist = np.append(dis_loss_hist, np.mean(discri_losses))
dis_loss_real_hist = np.append(dis_loss_real_hist,
np.mean(discri_loss_r))
dis_loss_fake_hist = np.append(dis_loss_fake_hist,
np.mean(discri_loss_f))
print(
"epoch: %d, [Dis_loss: %.3f (Dis_loss_real: %.3f, Dis_loss_fake: %.3f), Gen_loss: %.3f] "
% (epoch, np.mean(discri_losses), np.mean(discri_loss_r),
np.mean(discri_loss_f), np.mean(genera_losses)))
sample_images(
gen_model,
os.path.join(run_folder, "epochSamples", "sample-%d" % (epoch)),
noise)
gan_model.save_weights(
os.path.join(run_folder, "weights", "weights-%d.h5" % (epoch)))
epoch += 1
return [
gen_loss_hist, dis_loss_hist, dis_loss_real_hist, dis_loss_fake_hist
]
def run_training(train_dataset,
encoder,
decoder,
discriminator,
vaegan,
epoch_start,
epochs,
optimizers,
gamma,
noise,
run_folder,
label_smooth_r,
label_smooth_f,
history=None):
if epoch_start != 0:
enc_loss_hist, dec_loss_hist, dis_loss_hist = history[0], history[
1], history[2]
else:
enc_loss_hist, dec_loss_hist, dis_loss_hist = np.array([]), np.array(
[]), np.array([])
for epoch in range(epoch_start, epochs):
l = train(train_dataset, encoder, decoder, discriminator, optimizers,
gamma, label_smooth_r, label_smooth_f)
if (epoch + 1) % 1 == 0:
print('epoch:', epoch + 1, 'kl_loss:', l[0], 'rec_loss:', l[1],
'gen_loss:', l[2], 'dis_loss:', l[3])
enc_loss_hist = np.append(enc_loss_hist, (l[0] + l[1]))
dec_loss_hist = np.append(dec_loss_hist, (l[1] - l[3]))
dis_loss_hist = np.append(dis_loss_hist, l[3])
with open(os.path.join(run_folder, "history"), "wb") as fp:
history = np.array(
[enc_loss_hist, dec_loss_hist, dis_loss_hist])
pickle.dump(history, fp)
sample_images(
decoder,
os.path.join(run_folder, "epochSamples",
"sample-%d" % (epoch + 1)), noise, True)
vaegan.save_weights(
os.path.join(run_folder, "weights",
"weights-%d.h5" % (epoch + 1)))
return enc_loss_hist, dec_loss_hist, dis_loss_hist
def train(vae_model, decoder, train_set, validation_set, batch_size, no_epochs,
run_folder, noise):
""" Train the model and return loss and accuracy of train and validation set """
train_loss_hist = np.array([])
train_acc_hist = np.array([])
validation_loss_hist = np.array([])
validation_acc_hist = np.array([])
nb_batch = train_set.shape[0] // batch_size
idx_train = np.arange(len(train_set))
for epoch in range(1, no_epochs + 1):
np.random.shuffle(idx_train)
batch_losses = np.array([])
batch_acc = np.array([])
for i in range(nb_batch):
batch_x_train = train_set[idx_train[i * batch_size:(i + 1) *
batch_size]]
vae_loss, vae_acc = vae_model.train_on_batch(
batch_x_train, batch_x_train)
batch_losses = np.append(batch_losses, vae_loss)
batch_acc = np.append(batch_acc, vae_acc)
train_epoch_loss, train_epoch_acc = np.mean(batch_losses), np.mean(
batch_acc)
validation_loss, validation_acc = vae_model.evaluate(validation_set,
validation_set,
verbose=0)
train_loss_hist = np.append(train_loss_hist, train_epoch_loss)
train_acc_hist = np.append(train_acc_hist, train_epoch_acc)
validation_loss_hist = np.append(validation_loss_hist, validation_loss)
validation_acc_hist = np.append(validation_acc_hist, validation_acc)
print(
"epoch: %d, [TrainSet loss: %.3f, TrainSet accuracy: %.3f] [ValidSet loss: %.3f, ValidSet accuracy: %.3f] "
% (epoch, train_epoch_loss, train_epoch_acc, validation_loss,
validation_acc))
sample_images(
decoder,
os.path.join(run_folder, "epochSamples", "sample-%d" % (epoch)),
noise, True)
vae_model.save_weights(
os.path.join(run_folder, "weights", "weights-%d.h5" % (epoch)))
epoch += 1
return train_loss_hist, train_acc_hist, validation_loss_hist, validation_acc_hist
def plotting_process(dis, gen, gan, run_folder, latent_dim, w_filename):
# Load weights
gan.load_weights(w_filename)
# random latent space samples to plot
noise = np.random.normal(0, 1, (5 * 5, latent_dim))
sample_images(gen, os.path.join(run_folder, "sample-model"), noise)