def finetune(self, X):
print('Fine Tuning')
batches_per_epoch = int(len(X) / self.batch_size)
with tf.Session() as sess:
self.saver.restore(
sess,
tf.train.latest_checkpoint('./weights/sdae/' +
self.scope_name))
tmp = np.copy(X)
tmp = utils.add_noise(tmp, self.noise[0])
X = tmp
for epoch in range(self.epoch[0]):
avg_loss = 0.
self.learning_rate = utils.get_learning_rate(
self.learning_rate_decay, self.initial_learning_rate,
epoch)
for i in range(batches_per_epoch):
batch_x, batch_y = utils.get_batch(X, X, self.batch_size)
sess.run(self.finetuning_optimizer,
feed_dict={self.x[0]: batch_x})
loss = sess.run(self.finetuning_loss,
feed_dict={self.x[0]: batch_x})
avg_loss += loss
avg_loss /= batches_per_epoch
print('epoch {0}: loss = {1:.6f}'.format(epoch, avg_loss))
self.saver.save(sess,
'./weights/sdae/' + self.scope_name +
'/checkpoint',
global_step=0)
def train(self, X):
batches_per_epoch = int(len(X) / self.batch_size)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for layer in range(self.depth):
print('Layer {0}'.format(layer + 1))
tmp = np.copy(X)
tmp = utils.add_noise(tmp, self.noise[layer])
X = tmp
for epoch in range(self.epoch[layer]):
avg_loss = 0.
self.learning_rate = utils.get_learning_rate(
self.learning_rate_decay, self.initial_learning_rate,
epoch)
for i in range(batches_per_epoch):
batch_x, batch_y = utils.get_batch(
X, X, self.batch_size)
sess.run(self.layerwise_optimizers[layer],
feed_dict={self.x[layer]: batch_x})
loss = sess.run(self.layerwise_losses[layer],
feed_dict={self.x[layer]: batch_x})
avg_loss += loss
avg_loss /= batches_per_epoch
print("Epoch {0}: loss = {1:.6f}".format(epoch, avg_loss))
X = sess.run(self.layerwise_encoded[layer],
feed_dict={self.x[layer]: X})
self.saver.save(sess,
'./weights/sdae/' + self.scope_name +
'/checkpoint',
global_step=0)
def train(self, X, Y, lengths):
batches_per_epoch = int(len(X) / self.batch_size)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(self.epoch):
avg_loss = 0.
avg_accuracy = 0.
self.learning_rate = utils.get_learning_rate(self.learning_rate_decay, self.initial_learning_rate, epoch)
for i in range(batches_per_epoch):
batch_x, batch_y, batch_length = utils.get_rnn_sequential_batch(X, Y, lengths, i * self.batch_size, self.batch_size)
batch_x = utils.add_noise(batch_x, self.noise)
sess.run(self.optimizer, feed_dict={self.x: batch_x, self.y: batch_y, self.sequence_length: batch_length})
loss, accuracy = sess.run([self.loss, self.accuracy], feed_dict={self.x: batch_x, self.y: batch_y, self.sequence_length: batch_length})
avg_loss += loss
avg_accuracy += accuracy
avg_loss /= batches_per_epoch
avg_accuracy /= batches_per_epoch
print("Epoch {0}: loss = {1:.6f}, accuracy = {2:.2f}%".format(epoch, avg_loss, avg_accuracy * 100))
self.saver.save(sess, './weights/lstm/' + self.scope_name + '/checkpoint', global_step=0)
def train(self,
content_img_path='images/content/content1.jpg',
style_img_path='images/style/style1.jpg',
noise_img_path='images/content/content1.jpg',
output_img_path='results/anaoas',
tensorboard_path='tensorboard/tensorboard_anaoas',
show_img=None):
summ = tf.summary.merge_all()
with tf.Session() as sess:
ut = Utils()
noise_img_bytes = sess.run(
self.file_bytes, feed_dict={self.name_file: noise_img_path})
noise_img_np = np.fromstring(noise_img_bytes, np.uint8)
noise_img = np.reshape(
ut.add_noise(
ut.get_img(noise_img_np,
width=self.noise_img_width,
height=self.noise_img_height)),
(1, self.noise_img_height, self.noise_img_width,
self.noise_img_channels))
sess.run(tf.global_variables_initializer(),
feed_dict={self.noise_img_init: noise_img})
writer = tf.summary.FileWriter(tensorboard_path)
writer.add_graph(sess.graph)
content_img_bytes = sess.run(
self.file_bytes, feed_dict={self.name_file: content_img_path})
content_img_np = np.fromstring(content_img_bytes, np.uint8)
content_img = np.reshape(
ut.get_img(content_img_np,
width=self.content_img_width,
height=self.content_img_height),
(1, self.content_img_height, self.content_img_width,
self.content_img_channels))
style_img_bytes = sess.run(
self.file_bytes, feed_dict={self.name_file: style_img_path})
style_img_np = np.fromstring(style_img_bytes, np.uint8)
style_img = np.reshape(
ut.get_img(style_img_np,
width=self.style_img_width,
height=self.style_img_height),
(1, self.style_img_height, self.style_img_width,
self.style_img_channels))
for i in range(self.num_iters):
_, content_loss, style_loss, tv_loss, out_loss, out_img = sess.run(
[
self.optim, self.content_loss, self.style_loss,
self.total_variation_loss, self.total_loss,
self.noise_img
],
feed_dict={
self.content_img: content_img,
self.style_img: style_img
})
print('it: ', i)
print('Content loss: ', content_loss)
print('Style loss: ', style_loss)
print('Total variation loss: ', tv_loss)
print('Total loss: ', out_loss)
if i % 50 == 0:
decoded_img = ut.denormalize_img(out_img[0])
decoded_img = cv2.cvtColor(decoded_img, cv2.COLOR_BGR2RGB)
sess.run(self.fwrite,
feed_dict={
self.decoded_img:
decoded_img,
self.name_file:
output_img_path + '/img' + str(i) + '.png'
})
if show_img:
plt.axis("off")
plt.imshow(decoded_img)
plt.show()
s = sess.run(summ,
feed_dict={
self.content_img: content_img,
self.style_img: style_img
})
writer.add_summary(s, i)
