def z_iterator_visualize_model(self,
nr_iterations=1000,
plot_iteration_error=True):
print("Sampling images from model...")
batch_z = np.random.uniform(-1.0,
1.0,
size=[self.batch_size,
self.z_dim]).astype(np.float32)
feed_dict = {self.z_vec: batch_z, self.train_phase: False}
self.sess.run(self.init_z_iterator, feed_dict=feed_dict)
images = self.sess.run(self.gen_images_out,
feed_dict={self.train_phase: False})
#images = self.sess.run(self.gen_images, feed_dict=feed_dict)
images = utils.unprocess_image(images, 127.5, 127.5).astype(np.uint8)
shape = [4, self.batch_size // 4]
utils.save_imshow_grid(images,
self.logs_dir,
"generated_z_iterator.png",
shape=shape)
iterator_loss_store = []
for i in tqdm(xrange(nr_iterations)):
feed_dict = {self.train_phase: False}
_, iterator_loss = self.sess.run(
[self.z_iterator_train_op, self.gen_loss], feed_dict=feed_dict)
iterator_loss_store.append(iterator_loss)
if i == 0:
print("begining loss is " + str(iterator_loss))
print("final loss is " + str(iterator_loss))
images_iter = self.sess.run(self.gen_images_out, feed_dict=feed_dict)
images_iter = utils.unprocess_image(images_iter, 127.5,
127.5).astype(np.uint8)
print("diff is " + str(np.sum(images_iter - images)))
shape = [4, self.batch_size // 4]
utils.save_imshow_grid(images_iter,
self.logs_dir,
"generated_z_iterator_after.png",
shape=shape)
utils.save_imshow_grid(
np.abs(images_iter.astype(np.float) -
images.astype(np.float)).astype(np.uint8),
self.logs_dir,
"generated_z_iterator_dif.png",
shape=shape)
if plot_iteration_error:
fig = plt.figure(2)
iterator_loss_store = np.array(iterator_loss_store)
plt.plot(iterator_loss_store)
plt.xlabel("iteration step")
plt.ylabel("error (lower better)")
plt.title("Error vs z iteration step")
plt.show(fig)
def visualize_model(self):
print("Sampling images from model...")
batch_z = np.random.uniform(-1.0, 1.0, size=[self.batch_size, self.z_dim]).astype(np.float32)
feed_dict = {self.z_vec: batch_z, self.train_phase: False}
images = self.sess.run(self.gen_images, feed_dict=feed_dict)
images = utils.unprocess_image(images, 127.5, 127.5).astype(np.uint8)
shape = [4, self.batch_size // 4]
utils.save_imshow_grid(images, self.logs_dir, "generated.png", shape=shape)
def train_model(self, learning_rate, beta1, beta2, epsilon, max_iteration,
check_point):
with tf.Graph().as_default():
# initialize with random guess
logger.info(
'Initializing tensorflow graph with random guess......')
noise = np.random.normal(size=self.content_shape,
scale=np.std(self.content) * 0.1)
initial_guess = tf.random_normal(self.content_shape) * 0.256
input_image = tf.Variable(initial_guess)
parsed_net = self.vgg.load_net(input_image)
# calculate loss
content_loss = self._calculate_content_loss(parsed_net)
style_loss = self._calculate_style_loss(parsed_net)
tv_loss = self._calculate_tv_loss(input_image)
loss = content_loss + style_loss + tv_loss
# summary statistics
tf.summary.scalar('content_loss', content_loss)
tf.summary.scalar('style_loss', style_loss)
tf.summary.scalar('tv_loss', tv_loss)
tf.summary.scalar('total_loss', loss)
summary_loss = tf.summary.merge_all()
# initialize optimization
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
with tf.Session() as session:
summary_writer = tf.summary.FileWriter('logs/neural_network',
session.graph)
logger.info('Saving graph......')
session.run(tf.global_variables_initializer())
logger.info('Initializing optimization......')
logger.info('Current total loss: {}'.format(loss.eval()))
for k in range(max_iteration):
logger.info('Iteration {} total loss {}'.format(
str(k + 1), loss.eval()))
train_step.run()
summary = session.run(summary_loss)
summary_writer.add_summary(summary, k)
# save intermediate images at checkpoints
if (check_point and
(not k % check_point)) or k == max_iteration - 1:
output_temp = input_image.eval()
output_image = unprocess_image(
output_temp.reshape(self.content_shape[1:]),
self.vgg.mean_pixel)
yield k, output_image
def visualize_model(self, iterations, shape=[4, 16]):
print("Sampling images from model...")
batch_z = np.random.uniform(-1.0,
1.0,
size=[self.batch_size,
self.z_dim]).astype(np.float32)
feed_dict = {self.z_vec: batch_z, self.train_phase: False}
for cls in range(self.num_cls):
feed_dict[self.class_num] = cls
images = self.sess.run(self.gen_images, feed_dict=feed_dict)
#print(labels)
images = utils.unprocess_image(images, 127.5,
127.5).astype(np.uint8)
save_img_fn = "generated_cls" + str(cls) + "_" + str(
int(iterations)) + ".png"
utils.save_imshow_grid(images,
self.logs_dir,
save_img_fn,
shape=shape)
def visualize_model(self):
print("Sampling images from model...")
batch_z = np.random.uniform(-1.0,
1.0,
size=[self.batch_size,
self.z_dim]).astype(np.float32)
feed_dict = {self.z_vec: batch_z, self.train_phase: False}
images = self.sess.run(self.gen_images, feed_dict=feed_dict)
images = utils.unprocess_image(images, 127.5, 127.5).astype(np.uint8)
for i, image in enumerate(images):
Image.fromarray(image).save(
os.path.join(self.logs_dir, "%d_ge_testn.jpg" % i))
return
shape = [4, self.batch_size // 4]
utils.save_imshow_grid(images,
self.logs_dir,
"generated.png",
shape=shape)
def visualize_model(self, logdir=None):
if not logdir:
logdir = self.logs_dir
print(self.root_scope_name + "Sampling images from model...")
batch_z = np.random.uniform(-1.0,
1.0,
size=[self.batch_size,
self.z_dim]).astype(np.float32)
feed_dict = {self.z_vec: batch_z, self.train_phase: False}
images = self.sess.run(self.gen_images, feed_dict=feed_dict)
images = utils.unprocess_image(images, 127.5, 127.5).astype(np.uint8)
shape = [4, self.batch_size // 4]
utils.save_imshow_grid(images,
logdir,
"generated_palette.png",
shape=shape)
for i in range(len(images)):
scipy.misc.imsave(logdir + "/generated_image%d.png" % (i + 1),
images[i])
