def train_VAE():
model = VAE()
model.summary()
history = []
x_train = import_image(folder='train')
x_test = import_image(folder='test')
#x_train = import_image('test')
arbitrary = np.zeros([x_train.shape[0], 1024 * 2])
model.summary()
history = model.fit(x_train, [x_train, arbitrary],
epochs=200,
batch_size=128,
shuffle=True,
verbose=1,
validation_data=(x_test,
[x_test,
arbitrary[:x_test.shape[0]]]))
print("history =", history.history.keys())
output_loss1 = history.history['recons_loss']
output_loss2 = history.history['KLD_loss']
np.save('VAE2_recons_loss', output_loss1)
np.save('VAE2_KLD_loss', output_loss2)
model.save_weights('saved_model/VAE2_epochs60_weights.h5')
def problem3(model): # plot test image and calculate MSE
x_test = import_image(path=sys.argv[1])
decoded_imgs = model.predict(x_test)[0]
#output(decoded_imgs[:10], 'test')
pred = np.concatenate([x_test[:10], decoded_imgs], axis=0)
output_20(pred, sys.argv[2] + "fig1_3.jpg", 80)
get_MSE(x_test, decoded_imgs)
def problem5():
np.random.seed(0)
path = sys.argv[1]
x_test = import_image(sys.argv[1])
input_img = Input(shape=(64, 64, 3))
encoder_output = VAE_encoder(input_img)
encoder = Model(input_img, encoder_output)
weights_path = 'saved_model/VAE_weights.h5'
encoder.load_weights(weights_path, by_name=True)
h = encoder.predict(x_test)
print("h =", h.shape)
np.save('report/VAE_problem5_h', h)
def train_ACGAN(self):
self.x_train = import_image('train')
self.features = import_image_features('train')
#self.features_cat = to_categorical(self.features, 2)
print("feat =", self.features.shape, self.features[:5])
self.x_train = self.x_train * 2 - 1
print("x_train =", self.x_train.shape)
self.half_batch = int(self.batch_size / 2)
noise = np.random.normal(0, 1, (10, 128))
history = []
self.train_D()
self.train_G()
fout = open('process_ACGAN', 'w')
for epoch in range(self.epochs):
if self.d_loss[3] > 0.8: #g_loss[1] > 3:
print("%d" % 1, end=' ')
fout.write("%d " % 1)
for i in range(15):
self.train_G()
elif self.d_loss[3] > 0.65: #g_loss[1] > 1.5:
print("%d" % 2, end=' ')
fout.write("%d " % 2)
for i in range(5):
self.train_G()
elif self.d_loss[3] < 0.45: #[1] > 1.8:
print("%d" % 4, end=' ')
fout.write("%d " % 4)
for i in range(2):
self.train_D()
else:
print("%d" % 3, end=' ')
fout.write("%d " % 3)
self.train_D()
self.train_G()
'''
if self.g_loss[1] > 2: #g_loss[1] > 3:
print("%d" %1, end = ' ')
fout.write("%d " %1)
self.train_G()
elif self.g_loss[1] > 1.5: #g_loss[1] > 1.5:
print("%d" %2, end = ' ')
fout.write("%d " %2)
for i in range(5):
self.train_G()
self.train_D()
elif self.d_loss[1] > 2: #g_loss[1] > 1.5:
print("%d" %4, end = ' ')
fout.write("%d " %4)
self.train_D()
else:
print("%d" %3, end = ' ')
fout.write("%d " %3)
self.train_D()
self.train_G()
'''
# Plot the progress
if epoch % 10 == 0:
print("[acc1.: %.2f%%, acc2.: %.2f%%]" %
(100 * self.d_loss[3], 100 * self.d_loss[4])) # 3, 4
print("%d [D loss1: %.4f, D loss2.: %.4f] [G loss1: %.4f, G loss2: %.4f]" % \
(epoch, self.d_loss[1], self.d_loss[2], self.g_loss[1], self.g_loss[2])), # 1, 2
fout.write("%d [D loss: %f, acc.: %.2f%%] [G loss: %f]\n" %
(epoch, self.d_loss[0], 100 * self.d_loss[3],
self.g_loss[0]))
if epoch % self.loss_interval == 0:
history.append([self.d_loss[1], self.d_loss[2], self.g_loss[1], \
self.g_loss[2], 100*self.d_loss[3], 100*self.d_loss[4]])
# If at save interval => save generated image samples
if epoch % self.sample_interval == 0:
self.sample_images(epoch, noise)
if epoch % 1000 == 0:
self.generator.save_weights('saved_model/ACGAN_G_epochs' +
str(epoch) + '_weights.h5')
self.discriminator.save_weights('saved_model/ACGAN_D_epochs' +
str(epoch) + '_weights.h5')
self.combined.save_weights('saved_model/ACGAN_C_epochs' +
str(epoch) + '_weights.h5')
np.save("ACGAN_history", np.array(history))
#history.append([self.d_loss[0], 100*self.d_loss[1], self.g_loss])
self.generator.save_weights('saved_model/GAN_G_epochs' + str(epoch) +
'_weights.h5')
self.discriminator.save_weights('saved_model/GAN_D_epochs' +
str(epoch) + '_weights.h5')
self.combined.save_weights('saved_model/GAN_C_epochs' + str(epoch) +
'_weights.h5')
np.save("ACGAN_history", np.array(history))
self.sample_images(epoch, noise)
def train_GAN(self):
#self.x_train = import_image('test')
self.x_train = import_image('train')
#x_test = import_image('test')
self.x_train = self.x_train * 2 - 1
print("x_train =", self.x_train.shape)
self.half_batch = int(self.batch_size / 2)
print('construct noise')
noise = np.random.normal(0, 1, (32, 128))
np.save('GAN_noise.npy', noise)
history = []
self.train_D()
self.train_G()
fout = open('process2', 'a')
for epoch in range(9000, 9000 + self.epochs):
if self.d_loss[1] > 0.8:
print("%d" % 1, end=' ')
fout.write("%d " % 1)
for i in range(15):
self.train_G()
elif self.d_loss[1] > 0.65:
print("%d" % 2, end=' ')
fout.write("%d " % 2)
for i in range(5):
self.train_G()
elif self.d_loss[1] < 0.55:
print("%d" % 4, end=' ')
fout.write("%d " % 4)
for i in range(2):
self.train_D()
elif self.d_loss[1] < 0.4:
print("%d" % 5, end=' ')
fout.write("%d " % 5)
for i in range(7):
self.train_D()
else:
print("%d" % 3, end=' ')
fout.write("%d " % 3)
self.train_D()
self.train_G()
# Plot the progress
if epoch % 10 == 0:
print(
"%d [D loss: %f, acc.: %.2f%%] [G loss: %f]" %
(epoch, self.d_loss[0], 100 * self.d_loss[1], self.g_loss))
fout.write(
"%d [D loss: %f, acc.: %.2f%%] [G loss: %f]\n" %
(epoch, self.d_loss[0], 100 * self.d_loss[1], self.g_loss))
if epoch % self.loss_interval == 0:
history.append(
[self.d_loss[0], 100 * self.d_loss[1], self.g_loss])
# If at save interval => save generated image samples
if epoch % self.sample_interval == 0:
self.sample_images(epoch, noise)
if epoch % 500 == 0:
self.generator.save_weights('saved_model/GAN_G_epochs' +
str(epoch) + '_weights.h5')
self.discriminator.save_weights('saved_model/GAN_D_epochs' +
str(epoch) + '_weights.h5')
self.combined.save_weights('saved_model/GAN_C_epochs' +
str(epoch) + '_weights.h5')
np.save("GAN_history2", np.array(history))
history.append([self.d_loss[0], 100 * self.d_loss[1], self.g_loss])
self.generator.save_weights('saved_model/GAN_G_epochs' + str(epoch) +
'_weights.h5')
self.discriminator.save_weights('saved_model/GAN_D_epochs' +
str(epoch) + '_weights.h5')
self.combined.save_weights('saved_model/GAN_C_epochs' + str(epoch) +
'_weights.h5')
np.save("GAN_history2", np.array(history))
self.sample_images(20000, noise)