loss_discrim = -tf.reduce_sum(
discrim_target *
tf.log(tf.clip_by_value(discrim_predictions, 1e-10, 1.0))
) # Note: here use tf.reduce_sum, not use tf.reduce_mean
loss_texture = -loss_discrim
correct_predictions = tf.equal(tf.argmax(discrim_predictions, 1),
tf.argmax(discrim_target, 1))
discim_accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32))
# 2) content loss
CX_LAYER = 'conv4_2'
enhanced_vgg = vgg.net(vgg_dir, vgg.preprocess(enhanced * 255))
dslr_vgg = vgg.net(vgg_dir, vgg.preprocess(dslr_image * 255))
# SSIM loss
ssim_loss = 25 * (1 - utils.ssim(dslr_image, enhanced) / batch_size)
# CX loss
cx_loss = 4 * CX_loss_helper(dslr_vgg[CX_LAYER], enhanced_vgg[CX_LAYER],
config_CX)
# content loss
loss_content = ssim_loss + cx_loss
# 3) color loss
enhanced_blur = utils.blur(enhanced)
def stylize(network,
initial,
initial_noiseblend,
content,
styles,
preserve_colors,
iterations,
content_weight,
content_weight_blend,
style_weight,
style_layer_weight_exp,
style_blend_weights,
tv_weight,
learning_rate,
beta1,
beta2,
epsilon,
pooling,
exp_sigma,
mat_sigma,
mat_rho,
text_to_print,
print_iterations=None,
checkpoint_iterations=None,
kernel=3,
d=2,
gamma_rho=1,
gamma=1,
rational_rho=1,
alpha=1):
tf.logging.set_verbosity(tf.logging.INFO)
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
0 - dot product kernel
1 - exponential kernel
2 - matern kernel
3 - polynomial kernel
"""
shape = (1, ) + content.shape
style_shapes = [(1, ) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, pooling)
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
# sqr = features.T*features.T
# dim = features.shape
if (kernel == 0):
gram2 = np.matmul(features.T, features) / features.size
elif (kernel == 1):
gram2 = gramSquaredExp_np(
features,
exp_sigma) / features.size # exponential kernal
elif (kernel == 2):
gram2 = gramMatten_np(
features, mat_sigma, v,
mat_rho) / features.size # Mattern kernal
elif (kernel == 3):
print(d)
gram2 = gramPoly_np(features, C=0, d=d) / features.size
elif (kernel == 4):
gram2 = gramGammaExp_np(features, gamma_rho,
gamma) / features.size
elif (kernel == 5):
gram2 = gramRatioanlQuad_np(features, rational_rho,
alpha) / features.size
# print(features.shape,"diamention of feature\n")
style_features[i][layer] = gram2
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
g = tf.Graph()
with g.as_default(), g.device('/gpu'):
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (
tf.random_normal(shape) *
0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(
content_layers_weights[content_layer] * content_weight *
(2 * tf.nn.l2_loss(net[content_layer] -
content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value,
layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
style_gram = style_features[i][style_layer]
dim = feats.get_shape()
# print(dim)
sqr = tf.reduce_sum(tf.transpose(feats) * tf.transpose(feats),
axis=1)
d2 = tf.nn.relu(
tf.transpose(tf.ones([dim[1], dim[1]]) * sqr) +
tf.ones([dim[1], dim[1]]) * sqr -
2 * tf.matmul(tf.transpose(feats), feats))
if (kernel == 0):
gram = (tf.matmul(tf.transpose(feats), feats)) / size
elif (kernel == 1):
gram = tf.exp(
-1 * (tf.transpose(tf.ones([dim[1], dim[1]]) * sqr) +
tf.ones([dim[1], dim[1]]) * sqr -
2 * tf.matmul(tf.transpose(feats), feats)) / 2 /
(exp_sigma * exp_sigma)) / size # exponetial kernal
elif (kernel == 2):
# mattern kernal
if (v == 0.5):
gram = mat_sigma**2 * tf.exp(
-1 * tf.sqrt(d2) / mat_rho) / size
elif (v == 1.5):
gram = mat_sigma**2 * (tf.ones([
dim[1], dim[1]
]) + tf.sqrt(3.0) * tf.sqrt(d2) / mat_rho) * tf.exp(
-1 * tf.sqrt(3.0) * tf.sqrt(d2) / mat_rho) / size
elif (v == 2.5):
gram = mat_sigma**2 * (
tf.ones([dim[1], dim[1]]) +
tf.sqrt(5.0) * tf.sqrt(d2) / mat_rho + 5 * d2 / 3 /
(mat_rho**2)) * tf.exp(
-1 * tf.sqrt(5.0) * tf.sqrt(d2) /
mat_rho) / size
elif (kernel == 3):
# polynomial kernal
gram = (tf.matmul(tf.transpose(feats), feats))**d / size
elif (kernel == 4):
# gamma exponental kernal
gram = tf.exp(-1 * (tf.sqrt(d2) / gamma_rho)**gamma) / size
elif (kernel == 5):
# gamma exponental kernal
gram = (1 +
(d2 / rational_rho**2 / 2 / alpha))**(-1 *
alpha) / size
style_losses.append(style_layers_weights[style_layer] * 2 *
tf.nn.l2_loss(gram - style_gram) /
style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(
tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:, 1:, :, :])
tv_x_size = _tensor_size(image[:, :, 1:, :])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:, 1:, :, :] - image[:, :shape[1] - 1, :, :])
/ tv_y_size) +
(tf.nn.l2_loss(image[:, :, 1:, :] - image[:, :, :shape[2] - 1, :])
/ tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
# train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss)
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
def print_progress(last_loss):
new_loss = loss.eval()
stderr.write('file ===> %s \n' % text_to_print)
stderr.write(' content loss: %1.3e \t' % content_loss.eval())
stderr.write(' style loss: %1.3e \t' % style_loss.eval())
stderr.write(' tv loss: %1.3e \t' % tv_loss.eval())
stderr.write(' total loss: %1.3e \t' % new_loss)
stderr.write(' loss difference: %1.3e \t\n' %
(last_loss - new_loss))
return new_loss
def save_progress():
dict = {
"content loss": content_loss.eval(),
"style loss": style_loss.eval(),
"tv loss": tv_loss.eval(),
"total loss": loss.eval()
}
return dict
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
new_loss = 0
# if (print_iterations and print_iterations != 0):
# print_progress()
for i in range(iterations):
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations
and i % print_iterations == 0):
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
new_loss = print_progress(new_loss)
if (checkpoint_iterations
and i % checkpoint_iterations == 0) or last_step:
dict = save_progress()
this_loss = loss.eval()
print(this_loss, "loss in each check point")
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
try:
img_out = vgg.unprocess(best.reshape(shape[1:]),
vgg_mean_pixel)
except:
print(
"uanlabe to result image due to given parameters")
img_out = "no image"
if preserve_colors and preserve_colors:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(
Image.fromarray(
styled_grayscale_rgb.astype(
np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(
Image.fromarray(original_image.astype(
np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(
Image.fromarray(combined_yuv,
'YCbCr').convert('RGB'))
yield ((None if last_step else i), img_out, dict)
def train(self):
with tf.Session() as sess:
out_im = self.U_net(self.holder[41] / 127.5 - 1)
gt_resize = tf.image.resize_images(self.holder[42] / 127.5 - 1,
[256, 256])
image_pre = vgg.preprocess(gt_resize)
fai_imgt = {}
net = vgg.net(self.vgg_path, image_pre)
for layer in self.vgg_layer:
fai_imgt[layer] = net[layer]
image_pre = vgg.preprocess(
tf.image.resize_images(out_im, [256, 256]))
fai_imout = {}
net = vgg.net(self.vgg_path, image_pre)
for layer in self.vgg_layer:
fai_imout[layer] = net[layer]
Im_compt = self.holder[16] * self.holder[42] + (tf.add(
tf.multiply(self.holder[16], -1), 1)) * ((out_im + 1) * 127.5)
im_compt = tf.image.resize_images(Im_compt / 127.5 - 1, [256, 256])
image_pre = vgg.preprocess(im_compt)
fai_compt = {}
net = vgg.net(self.vgg_path, image_pre)
for layer in self.vgg_layer:
fai_compt[layer] = net[layer]
U_vars = [
var for var in tf.trainable_variables() if 'UNET' in var.name
]
total_loss = get_total_loss(out_im, self.holder[-1] / 127.5 - 1,
self.holder[16], fai_imout, fai_imgt,
fai_compt, self.vgg_layer, im_compt)
optim = tf.train.AdamOptimizer()
optimizer = optim.minimize(total_loss[0], var_list=U_vars)
int_group = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(int_group)
graph = tf.summary.FileWriter(self.logdir, sess.graph)
saver = tf.train.Saver(U_vars, max_to_keep=20)
for epoch in range(self.num_epochs):
for imid in range(int(self.total_ims // self.batch)):
mask_ims, gt_ims = get_im(self.ims_dir, imid)
self.get_all_mask(mask_ims, gt_ims)
feed_dic = get_feedict(self.all_masks, self.holder)
_, loss_total = sess.run([optimizer, total_loss],
feed_dict=feed_dic)
if (int(epoch * self.total_ims) + imid) % 1 == 0:
print(
'epoch: %d, cur_num: %d, total_loss: %f, l_hole: %f, l_valid: %f, percept_loss: %f, style_loss_out: %f, style_loss_comp: %f, tv_loss: %f'
% (epoch, imid, loss_total[0], loss_total[1],
loss_total[2], loss_total[3], loss_total[4],
loss_total[5], loss_total[6]))
if epoch % 5 == 0:
saver.save(sess,
self.save_path + 'model.ckpt',
global_step=epoch)
def stylize(network, initial, content, styles, iterations,
content_weight, style_weight, style_blend_weights, tv_weight,
learning_rate, print_iterations=None, checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/gpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net, mean_pixel = vgg.net(network, image)
content_pre = np.array([vgg.preprocess(content, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/gpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net, _ = vgg.net(network, image)
style_pre = np.array([vgg.preprocess(styles[i], mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, mean_pixel)])
initial = initial.astype('float32')
image = tf.Variable(initial)
net, _ = vgg.net(network, image)
# content loss
content_loss = content_weight * (2 * tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) /
content_features[CONTENT_LAYER].size)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
def print_progress(i, last=False):
stderr.write('Iteration %d/%d\n' % (i + 1, iterations))
if last or (print_iterations and i % print_iterations == 0):
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(iterations):
last_step = (i == iterations - 1)
print_progress(i, last=last_step)
train_step.run()
if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
yield (
(None if last_step else i),
vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
)
def optimize(content_targets,
style_targets,
content_weight,
style_weight,
tv_weight,
vgg_path,
epochs=2,
print_iterations=1,
batch_size=4,
save_path='saver/fns.ckpt',
slow=False,
learning_rate=1e-3,
debug=False,
save_checkpoint=False,
restore_checkpoint_path=None):
if slow:
batch_size = 1
mod = len(content_targets) % batch_size
if mod > 0:
print("Train set has been trimmed slightly..")
content_targets = content_targets[:-mod]
# style_features = collections.defaultdict()
style_features = []
batch_shape = (batch_size, 256, 256, 5)
style_shape = (1, 256, 256, 3)
# print(style_shape)
# precompute style features
with tf.Graph().as_default(), tf.device('/cpu:0'), tf.Session() as sess:
# with tf.Graph().as_default(), tf.Session() as sess:
style_image = tf.placeholder(tf.float32,
shape=style_shape,
name='style_image')
style_image_pre = vgg.preprocess(style_image)
net = vgg.net(vgg_path, style_image_pre)
for i in range(len(style_targets)):
index = 0
style_pre = np.array([style_targets[i]])
# current_style_feature = []
current_style_feature = np.array([])
for layer in STYLE_LAYERS:
# print(layer)
features = net[layer].eval(feed_dict={style_image: style_pre})
# print(features.shape)
features = np.reshape(features, (-1, features.shape[3]))
# print(features.shape)
# gram = np.matmul(features.T, features) / features.size
gram = np.matmul(features.T - np.mean(features.T),
features - np.mean(features)) / features.size
# print(gram.shape)
if not STYLE_LAYERS_SHAPE[STYLE_LAYERS.index(layer)]:
STYLE_LAYERS_SHAPE[STYLE_LAYERS.index(layer)] = gram.shape
if not STYLE_LAYERS_SIZE[STYLE_LAYERS.index(layer)]:
STYLE_LAYERS_SIZE[STYLE_LAYERS.index(layer)] = gram.size
if STYLE_LAYERS_INDEX[STYLE_LAYERS.index(layer)] == -1:
STYLE_LAYERS_INDEX[STYLE_LAYERS.index(layer)] = index
index = index + gram.size
# style_features[i][layer] = gram
# current_style_feature.append(gram.tolist())
# current_style_feature.append(gram.reshape(-1))
current_style_feature = np.append(current_style_feature,
gram.reshape(-1))
# style_features.append(np.array(current_style_feature).reshape(-1))
style_features.append(current_style_feature)
style_features = np.array(style_features, dtype=np.float32)
# tf.convert_to_tensor(style_features)
with tf.Graph().as_default(), tf.Session() as sess:
lambda_style = tf.placeholder(tf.float32, name="lambda_style")
style_id = tf.placeholder(tf.int32, name="style_id")
X_content = tf.placeholder(tf.float32,
shape=batch_shape,
name="X_content")
X_pre = vgg.preprocess(X_content[:, :, :, 0:3])
# precompute content features
content_features = {}
content_net = vgg.net(vgg_path, X_pre)
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER]
if slow:
preds = tf.Variable(
tf.random_normal(X_content.get_shape()) * 0.256)
preds_pre = preds
else:
preds = transform.net(X_content / 255.0)
preds_pre = vgg.preprocess(preds)
net = vgg.net(vgg_path, preds_pre)
content_size = _tensor_size(
content_features[CONTENT_LAYER]) * batch_size
assert _tensor_size(content_features[CONTENT_LAYER]) == _tensor_size(
net[CONTENT_LAYER])
# content_loss = (1 - lambda_style) * (2 * tf.nn.l2_loss(
# net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) / content_size
# )
content_loss = content_weight * (
2 * tf.nn.l2_loss(net[CONTENT_LAYER] -
content_features[CONTENT_LAYER]) / content_size
) # original
style_losses = []
for style_layer in STYLE_LAYERS:
# print(style_layer)
layer = net[style_layer]
bs, height, width, filters = map(lambda i: i.value,
layer.get_shape())
size = height * width * filters
feats = tf.reshape(layer, (bs, height * width, filters))
feats_T = tf.transpose(feats, perm=[0, 2, 1])
# grams = tf.matmul(feats_T, feats) / size
grams = tf.matmul(feats_T - tf.reduce_mean(feats_T),
feats - tf.reduce_mean(feats)) / size
# test = lambda_style.eval(session=sess)
# print('test : ' + test)
# style_gram = style_features[style_id.eval(session=sess)][style_layer]
# style_gram = style_features[0][style_layer]
# s_id = sess.run(style_id)
# style_gram = style_features[0][style_layer]
# style_features = [tf.convert_to_tensor(x) for x in style_features]
# s_gram = tf.gather_nd(style_features, style_id)
# style_gram = s_gram[style_layer]
# make style_features into tensor
# keys = []
# values = []
#
# for k, v in style_features.items():
# keys.append(k)
# values.append(k)
# style_features_tf = tf.contrib.lookup.HashTable(
# initializer=tf.contrib.lookup.KeyValueTensorInitializer(
# keys=tf.constant(range(len(style_targets))),
# # values=tf.constant([style_features[i] for i in range(len(style_targets))]),
# # values=tf.constant(style_targets),
# values=style_targets,
# ),
# default_value=tf.constant(-1),
# name="style_features_tf"
# )
# current_style_features = style_features_tf.lookup(style_id)
# style_gram = tf.gather_nd(current_style_features, STYLE_LAYERS.index(style_layer))
# style_features_tf.init.run()
# print(style_gram.eval())
# print(style_gram.eval())
style_index = STYLE_LAYERS.index(style_layer)
style_grams = tf.gather_nd(tf.constant(style_features), [style_id])
style_gram = style_grams[STYLE_LAYERS_INDEX[style_index]:
STYLE_LAYERS_INDEX[style_index] +
STYLE_LAYERS_SIZE[style_index]]
style_gram = tf.reshape(style_gram,
STYLE_LAYERS_SHAPE[style_index])
# style_grams = style_features[style_id]
# style_losses.append(2 * tf.nn.l2_loss(grams - style_gram)/style_gram.size)
style_losses.append(2 * tf.nn.l2_loss(grams - style_gram) /
STYLE_LAYERS_SIZE[style_index])
style_loss = lambda_style * functools.reduce(tf.add,
style_losses) / batch_size
# style_loss = style_weight * functools.reduce(tf.add, style_losses) / batch_size # original
# total variation denoising
tv_y_size = _tensor_size(preds[:, 1:, :, :])
tv_x_size = _tensor_size(preds[:, :, 1:, :])
y_tv = tf.nn.l2_loss(preds[:, 1:, :, :] -
preds[:, :batch_shape[1] - 1, :, :])
x_tv = tf.nn.l2_loss(preds[:, :, 1:, :] -
preds[:, :, :batch_shape[2] - 1, :])
tv_loss = tv_weight * 2 * (x_tv / tv_x_size +
y_tv / tv_y_size) / batch_size
loss = content_loss + style_loss + tv_loss
# overall loss
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
sess.run(tf.global_variables_initializer())
import random
uid = random.randint(1, 100)
print("UID: %s" % uid)
if restore_checkpoint_path:
print('Restoring checkpoint from : ' + restore_checkpoint_path)
saver = tf.train.Saver()
saver.restore(sess, restore_checkpoint_path)
for epoch in range(epochs):
if save_checkpoint and epoch > 0:
saver = tf.train.Saver()
head, tail = os.path.split(save_path)
cp_dir = os.path.join(head, str(epoch))
if not os.path.exists(cp_dir):
os.makedirs(cp_dir)
cp_path = os.path.join(cp_dir, tail)
print('save checkpoint to : ' + cp_path)
res = saver.save(sess, cp_path)
print('epoch: {}'.format(epoch))
num_examples = len(content_targets)
iterations = 0
while iterations * batch_size < num_examples:
# print('iterations: {}'.format(iterations))
start_time = time.time()
curr = iterations * batch_size
curr_lambda_style = np.random.randint(1, 100) * 1.0
curr_lambda_style_img = np.ones(
(256, 256, 1)) * curr_lambda_style
curr_style_id = np.random.randint(
len(style_targets)) if epoch > 0 else 0
# print('\ncurr_style_id:')
# print(type(curr_style_id))
# print(curr_style_id)
curr_style_channel = np.ones((256, 256, 1)) * curr_style_id
step = curr + batch_size
X_batch = np.zeros(batch_shape, dtype=np.float32)
for j, img_p in enumerate(content_targets[curr:step]):
try:
curr_img = get_img(img_p,
(256, 256, 3)).astype(np.float32)
except Exception:
continue
X_batch[j, :, :, 0:3] = curr_img
X_batch[j, :, :, 3:] = curr_lambda_style_img
X_batch[j, :, :, 4:] = curr_style_channel
iterations += 1
assert X_batch.shape[0] == batch_size
feed_dict = {
X_content: X_batch,
lambda_style: curr_lambda_style,
style_id: curr_style_id
}
train_step.run(feed_dict=feed_dict)
end_time = time.time()
delta_time = end_time - start_time
if debug:
print("UID: %s, batch time: %s" % (uid, delta_time))
is_print_iter = int(iterations) % print_iterations == 0
if slow:
is_print_iter = epoch % print_iterations == 0
is_last = epoch == epochs - 1 and iterations * batch_size >= num_examples
should_print = is_print_iter or is_last
if should_print:
to_get = [style_loss, content_loss, tv_loss, loss, preds]
test_feed_dict = {
X_content: X_batch,
lambda_style: 80.0,
style_id: 0 # np.random.randint(1, 10) / 10.0
}
tup = sess.run(to_get, feed_dict=test_feed_dict)
_style_loss, _content_loss, _tv_loss, _loss, _preds = tup
losses = (_style_loss, _content_loss, _tv_loss, _loss)
# print(losses)
if slow:
_preds = vgg.unprocess(_preds)
else:
saver = tf.train.Saver()
res = saver.save(sess, save_path)
yield (_preds, losses, iterations, epoch)
def main():
parser = build_parser()
options, unknown = parser.parse_known_args()
env = os.environ.copy()
print("options: ", options)
vgg_path = options.dataset + '/vgg/imagenet-vgg-verydeep-19.mat'
model_name = options.style_image.replace('.jpg', '.ckpt')
style_image = options.dataset + '/style_images/' + options.style_image
training_path = options.dataset + '/train'
model_dir = env.get("OUTPUT_DIR", options.ckpt)
tensorboard_dir = env.get("LOG_DIR", options.dataset)
print("style_image: ", style_image)
print("vgg: ", vgg_path)
print("trainingpath: ", training_path)
print("modelname: ", model_name)
if options.gpu == None:
available_gpus = get_available_gpus()
if len(available_gpus) > 0:
device = '/gpu:0'
else:
device = '/cpu:0'
else:
if options.gpu > -1:
device = '/gpu:{}'.format(options.gpu)
else:
device = '/cpu:0'
batchsize = options.batchsize
# content targets
content_targets = [os.path.join(training_path, fn) for fn in list_files(training_path)]
if len(content_targets) % batchsize != 0:
content_targets = content_targets[:-(len(content_targets) % batchsize)]
print('total training data size: ', len(content_targets))
batch_shape = (batchsize,224,224,3)
# style target
style_target = read_img(style_image)
style_shape = (1,) + style_target.shape
with tf.device(device), tf.Session() as sess:
# style target feature
# compute gram maxtrix of style target
style_image = tf.placeholder(tf.float32, shape=style_shape, name='style_image')
vggstyletarget = vgg.net(vgg_path, vgg.preprocess(style_image))
style_vgg = vgg.get_style_vgg(vggstyletarget, style_image, np.array([style_target]))
# content target feature
content_vgg = {}
inputs = tf.placeholder(tf.float32, shape=batch_shape, name="inputs")
content_net = vgg.net(vgg_path, vgg.preprocess(inputs))
content_vgg['relu4_2'] = content_net['relu4_2']
# feature after transformation
outputs = stylenet.net(inputs/255.0)
vggoutputs = vgg.net(vgg_path, vgg.preprocess(outputs))
# compute feature loss
loss_f = options.lambda_feat * vgg.total_content_loss(vggoutputs, content_vgg, batchsize)
# compute style loss
loss_s = options.lambda_style * vgg.total_style_loss(vggoutputs, style_vgg, batchsize)
# total variation denoising
loss_tv = options.lambda_tv * vgg.total_variation_regularization(outputs, batchsize, batch_shape)
# total loss
loss = loss_f + loss_s + loss_tv
with tf.Session() as sess:
if not os.path.exists(options.ckpt):
os.makedirs(options.ckpt)
save_path = model_dir + '/' + model_name
# op to write logs to Tensorboard
#training
train_step = tf.train.AdamOptimizer(options.lr).minimize(loss)
sess.run(tf.global_variables_initializer())
total_step = 0
for epoch in range(options.epoch):
print('epoch: ', epoch)
step = 0
while step * batchsize < len(content_targets):
time_start = time.time()
batch = np.zeros(batch_shape, dtype=np.float32)
for i, img in enumerate(content_targets[step * batchsize : (step + 1) * batchsize]):
batch[i] = read_img(img).astype(np.float32) # (224,224,3)
step += 1
total_step += 1
loss_, summary= sess.run([loss, train_step,], feed_dict= {inputs:batch})
time_elapse = time.time() - time_start
should_save = total_step % 2000 == 0
if total_step % 1 == 0:
print('[step {}] elapse time: {} loss: {}'.format(total_step, time_elapse, loss_))
if should_save:
print('Saving checkpoint')
saver = tf.train.Saver()
res = saver.save(sess, save_path)
print('Saving final result to ' + save_path)
saver = tf.train.Saver()
res = saver.save(sess, save_path)
def stylyze(options, callback):
parser = build_parser()
if options is None:
key = 'TF_CPP_MIN_LOG_LEVEL'
if key not in os.environ:
os.environ[key] = '2'
options = parser.parse_args()
if not os.path.isfile(options.network):
parser.error("Network %s does not exist. (Did you forget to "
"download it?)" % options.network)
if [options.checkpoint_iterations,
options.checkpoint_output].count(None) == 1:
parser.error("use either both of checkpoint_output and "
"checkpoint_iterations or neither")
if options.checkpoint_output is not None:
if re.match(r'^.*(\{.*\}|%.*).*$', options.checkpoint_output) is None:
parser.error("To save intermediate images, the checkpoint_output "
"parameter must contain placeholders (e.g. "
"`foo_{}.jpg` or `foo_%d.jpg`")
content_image_arr = [imread(i) for i in options.content]
style_images = [imread(style) for style in options.styles]
width_arr = options.width
for i in range(len(content_image_arr)):
width = width_arr[i]
content_image = content_image_arr[i]
if width is not None:
new_shape = (int(
math.floor(
float(content_image.shape[0]) / content_image.shape[1] *
width)), width)
content_image = scipy.misc.imresize(content_image, new_shape)
content_image_arr[i] = content_image
target_shape = content_image.shape
for j in range(len(style_images)):
style_scale = STYLE_SCALE
if options.style_scales is not None:
style_scale = options.style_scales[j]
style_images[j] = scipy.misc.imresize(
style_images[j],
style_scale * target_shape[1] / style_images[j].shape[1])
style_blend_weights = options.style_blend_weights
if style_blend_weights is None:
# default is equal weights
style_blend_weights = [1.0 / len(style_images) for _ in style_images]
else:
total_blend_weight = sum(style_blend_weights)
style_blend_weights = [
weight / total_blend_weight for weight in style_blend_weights
]
initial_arr = content_image_arr
# try saving a dummy image to the output path to make sure that it's writable
output_arr = options.output
for output in output_arr:
if os.path.isfile(output) and not options.overwrite:
raise IOError("%s already exists, will not replace it without "
"the '--overwrite' flag" % output)
try:
imsave(output, np.zeros((500, 500, 3)))
except:
raise IOError('%s is not writable or does not have a valid file '
'extension for an image file' % output)
vgg_weights, vgg_mean_pixel = vgg.load_net(options.network)
style_shapes = [(1, ) + style.shape for style in style_images]
style_features = [{} for _ in style_images]
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= options.style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute style features in feedforward mode
for i in range(len(style_images)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, options.pooling)
style_pre = np.array(
[vgg.preprocess(style_images[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - options.initial_noiseblend
for i in range(len(content_image_arr)):
Data.save_step(Data.get_step() + 1)
loss_arrs = None
for iteration, image, loss_vals in stylize(
initial=initial_arr[i],
content=content_image_arr[i],
preserve_colors=options.preserve_colors,
iterations=options.iterations,
content_weight=options.content_weight,
content_weight_blend=options.content_weight_blend,
tv_weight=options.tv_weight,
learning_rate=options.learning_rate,
beta1=options.beta1,
beta2=options.beta2,
epsilon=options.epsilon,
pooling=options.pooling,
initial_content_noise_coeff=initial_content_noise_coeff,
style_images=style_images,
style_layers_weights=style_layers_weights,
style_weight=options.style_weight,
style_blend_weights=style_blend_weights,
vgg_weights=vgg_weights,
vgg_mean_pixel=vgg_mean_pixel,
style_features=style_features,
print_iterations=options.print_iterations,
checkpoint_iterations=options.checkpoint_iterations,
callback=callback):
if (image is not None) and (options.checkpoint_output is not None):
imsave(fmt_imsave(options.checkpoint_output, iteration), image)
if (loss_vals is not None) \
and (options.progress_plot or options.progress_write):
if loss_arrs is None:
itr = []
loss_arrs = OrderedDict(
(key, []) for key in loss_vals.keys())
for key, val in loss_vals.items():
loss_arrs[key].append(val)
itr.append(iteration)
imsave(options.output[i], image)
if options.progress_write:
fn = "{}/progress.txt".format(os.path.dirname(options.output[i]))
tmp = np.empty((len(itr), len(loss_arrs) + 1), dtype=float)
tmp[:, 0] = np.array(itr)
for ii, val in enumerate(loss_arrs.values()):
tmp[:, ii + 1] = np.array(val)
np.savetxt(fn,
tmp,
header=' '.join(['itr'] + list(loss_arrs.keys())))
if options.progress_plot:
import matplotlib
matplotlib.use('Agg')
from matplotlib import pyplot as plt
fig, ax = plt.subplots()
for key, val in loss_arrs.items():
ax.semilogy(itr, val, label=key)
ax.legend()
ax.set_xlabel("iterations")
ax.set_ylabel("loss")
fig.savefig("{}/progress.png".format(
os.path.dirname(options.output[i])))
def stylize(network, initial, content, styles, iterations,
content_weight, style_weight, style_blend_weights, tv_weight,
learning_rate, print_iterations=None, checkpoint_iterations=None,
print_image_iterations=False):
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net, mean_pixel = vgg.net(network, image)
content_pre = np.array([vgg.preprocess(content, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net, _ = vgg.net(network, image)
style_pre = np.array([vgg.preprocess(styles[i], mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, mean_pixel)])
initial = initial.astype('float32')
image = tf.Variable(initial)
net, _ = vgg.net(network, image)
# content loss
content_loss = content_weight * (2 * tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) /
content_features[CONTENT_LAYER].size)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
def print_progress(i, last=False):
if print_iterations is not None:
if i is not None and i % print_iterations == 0 or last:
print >> stderr, ' content loss: %g' % content_loss.eval()
print >> stderr, ' style loss: %g' % style_loss.eval()
print >> stderr, ' tv loss: %g' % tv_loss.eval()
print >> stderr, ' total loss: %g' % loss.eval()
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(iterations):
print_progress(i)
print >> stderr, 'Iteration %d/%d' % (i + 1, iterations)
train_step.run()
if (checkpoint_iterations is not None and
i % checkpoint_iterations == 0) or i == iterations - 1:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
print_progress(None, i == iterations - 1)
if (i % 100 == 0) and (print_image_iterations):
temp_image = vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
temp_output = 'iteration_' + str(i) + '.jpg'
imsave(temp_output, temp_image)
return vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
def stylize(network,
initial,
initial_noiseblend,
content,
styles,
preserve_colors,
iterations,
content_weight,
content_weight_blend,
style_weight,
style_layer_weight_exp,
style_blend_weights,
tv_weight,
learning_rate,
beta1,
beta2,
epsilon,
pooling,
print_iterations=None,
checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1, ) + content.shape
style_shapes = [(1, ) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
g = tf.Graph()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = '0'
with g.as_default(), g.device('/cpu:0'), tf.Session(config=config) as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session(
config=config) as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, pooling)
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (
tf.random_normal(shape) *
0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
'''
Compute the content loss
Variables:
content_weight: scalar constant we multiply the content_loss by.
net[content_layer]: features of the current image, Tensor with shape [1, height, width, channels]
content_features[content_layer]: features of the content image, Tensor with shape [1, height, width, channels]
'''
# features of the current image [1, height, width, channels]
l_content = content_weight * tf.reduce_sum(
(net[content_layer] - content_features[content_layer])**2)
content_losses.append(content_layers_weights[content_layer] *
l_content)
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, channels = map(lambda i: i.value,
layer.get_shape())
size = height * width * channels
'''
Compute the Gram matrix of the layer
Variables:
layer: features of the current image at style_layer, Tensor with shape [1, height, width, channels]
gram: computed gram matrix with shape [channels, channels]
'''
feats = tf.reshape(layer, (-1, channels))
gram = tf.matmul(tf.transpose(feats), feats)
gram /= size
'''
Compute the style loss
Variables:
style_layers_weights[style_layer]: scalar constant we multiply the content_loss by.
gram: computed gram matrix with shape [channels, channels]
style_gram: computed gram matrix of the style image at style_layer with shape [channels, channels]
'''
style_gram = style_features[i][style_layer]
l_style = style_layers_weights[style_layer] * tf.reduce_sum(
(gram - style_gram)**2)
style_losses.append(l_style)
style_loss += style_weight * style_blend_weights[i] * reduce(
tf.add, style_losses)
# total variation denoising
'''
Compute the TV loss
Variables:
tv_weight: scalar giving the weight to use for the TV loss.
image: tensor of shape (1, H, W, 3) holding current image.
'''
tv_loss = tv_weight * (tf.reduce_sum(
(image[:, 1:, :, :] - image[:, :-1, :, :])**2) + tf.reduce_sum(
(image[:, :, 1:, :] - image[:, :, :-1, :])**2))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in range(iterations):
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations
and i % print_iterations == 0):
print_progress()
if (checkpoint_iterations
and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]),
vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(
Image.fromarray(
styled_grayscale_rgb.astype(
np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(
Image.fromarray(original_image.astype(
np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(
Image.fromarray(combined_yuv,
'YCbCr').convert('RGB'))
yield ((None if last_step else i), img_out)
def stylize(network, initial, initial_noiseblend, content, styles, preserve_colors, iterations,
content_weight, content_weight_blend, style_weight, style_layer_weight_exp, style_blend_weights, tv_weight,
learning_rate, beta1, beta2, epsilon, pooling,
print_iterations=None, checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
# LOAD WEIGHTS AND AVG PIXEL HERE
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
# SET UP GRAPH
# RUN SESSION ON CPU
# IMAGE PLACEHOLDER
# LOAD VGG
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
# evaluate features
content_features[layer] = net[layer].eval(feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
# CREATE GRAPH FOR EVERY STYLE
# RUN SESSION ON CPU
# IMAGE PLACEHOLDER
# LOAD VGG
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
# evaluate features
features = net[layer].eval(feed_dict={image: style_pre})
# create gram matrix
features = np.reshape(features, (-1, features.shape[3]))
# CREATE GRAM MATRIX
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropagation
# USE DEFAULT GRAPH
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (tf.random_normal(shape) * 0.256) * (1.0 - initial_content_noise_coeff)
# CREATE IMAGE VARIABLE FROM INITIAL
# LOAD NET
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(content_layers_weights[content_layer] * content_weight * (2 * tf.nn.l2_loss(
net[content_layer] - content_features[content_layer]) /
content_features[content_layer].size))
# ADD CONTENT LOSSES FOR EACH LAYER TOGETHER - USE REDUCE
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
# CREATE GRAM MATRIX
style_gram = style_features[i][style_layer]
style_losses.append(style_layers_weights[style_layer] * 2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
# LOSS FUNCTION LINE HERE
# optimizer setup
# TF.TRAIN LINE HERE - USE ADAM
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization BACKPROP
best_loss = float('inf')
best = None
# START SESSION
# INITIALIZE VARIABLES TO BEGIN OPTIMIZATION
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
iteration_times = []
start = time.time()
for i in range(iterations):
iteration_start = time.time()
if i > 0:
elapsed = time.time() - start
# take average of last couple steps to get time per iteration
remaining = np.mean(iteration_times[-10:]) * (iterations - i)
stderr.write('Iteration %4d/%4d (%s elapsed, %s remaining)\n' % (
i + 1,
iterations,
hms(elapsed),
hms(remaining)
))
else:
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations and i % print_iterations == 0):
print_progress()
# backprop - replacing loss
if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
# CHECK LOSS HERE
img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(Image.fromarray(original_image.astype(np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(Image.fromarray(combined_yuv, 'YCbCr').convert('RGB'))
yield (
(None if last_step else i),
img_out
)
iteration_end = time.time()
iteration_times.append(iteration_end - iteration_start)
def stylize(network, initial, content, styles, iterations, content_weight, style_weight, style_blend_weights, tv_weight, learning_rate, print_iterations=None, checkpoint_iterations=None):
# input.shape = (n_image, height, width, channel)
content_shape = (1, ) + content.shape
# style.shape = [ all style shapes ]
style_shapes = [(1, ) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles] # for multiple style image inputs
# compute content features
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
content_pl = tf.placeholder('float', shape=content_shape)
# compute feedforward activations
# net is the activation of image using Placeholder
activation, mean_pixel = vgg.net(network, content_pl)
# preprocessing the input
content_preprocessed = np.array([vgg.preprocess(content, mean_pixel)])
# extract content features using preprocessed input into the VGG
# we only extract content features from one layer
content_features[CONTENT_LAYER] = activation[CONTENT_LAYER].eval(
feed_dict={content_pl: content_preprocessed})
# compute style features
# the loop below is for multiple style image inputs
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
# since different style layers have differnt shapes
style_pl = tf.placeholder('float', shape=style_shapes[i])
# question: why we use the mean value from content_features
activation, _ = vgg.net(network, style_pl)
style_preprocessed = np.array([vgg.preprocess(styles[i], mean_pixel)])
# since we will compute multiple layers of styles, we use loop
for layer in STYLE_LAYERS:
# extract the one of the style features from one layer
_features = activation[layer].eval(feed_dict={style_pl: style_preprocessed})
# since we will compute the Gram Matrix, we will reshape the output
# so that the inner product is easier to compute
# question why should we reshape? what is the origal shape? what does -1 mean
_features = _features.reshape((-1, _features.shape[3]))
# compute the Gram Matrix as style features
# why divide it by _features.size
gram = np.matmul(_features.T, _features) / _features.size
style_features[i][layer] = gram # the first index is the n_th style image input
# compute back-prop
with tf.Graph().as_default():
# initial = None means this iteration is our first iteration
# thus we need to generate a white noise image
if initial is None:
# the noise turned out to be not used at all
white_noise_image = np.random.normal(size=content_shape, scale=np.std(content) * .1)
initial = tf.random_normal(content_shape) * .256
# if we already have an image in training
# we will keep using this image for further modification
else:
initial_preprocessed = np.array([vgg.preprocess(initial, mean_pixel)])
initial = initial_preprocessed.astype('float32')
# we make this initial input as a trainable variable
image = tf.Variable(initial)
activation, _ = vgg.net(network, image)
# compute content loss
image_content_features = activation[CONTENT_LAYER]
target_content_features = content_features[CONTENT_LAYER]
# why divide it by target.size, can we eliminate that?
# the content weight is included here rather than the end
content_loss = content_weight * .5 * 1 / target_content_features.size * tf.nn.l2_loss(image_content_features - target_content_features)
# compute style loss
# using loop to sum style loss for multiple style image inputs
style_loss_for_all_styles = 0
for i in range(len(styles)):
style_losses = [] # the total losses
# using loop to sum style loss for multiple style layers
for style_layer in STYLE_LAYERS:
layer_activation = activation[style_layer]
_, height, width, channel = map(
lambda i: i.value, layer_activation.get_shape())
layer_size = height * width * channel
feats = tf.reshape(layer_activation, (-1, channel))
# it doesn't have to divide by size
image_style_gram = tf.matmul(tf.transpose(feats), feats) / layer_size
target_style_gram = style_features[i][style_layer]
layer_style_loss = 2 / target_style_gram.size * tf.nn.l2_loss(image_style_gram - target_style_gram)
style_losses.append(layer_style_loss)
style_loss_for_all_styles += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
# this loss is added to regularize that
# the output image will not deviate too much
# from content image at each pixel
tv_y_size = _tensor_size(image[:, 1:, :, :])
tv_x_size = _tensor_size(image[:, :, 1:, :])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:, 1:, :, :] - image[:, :content_shape[1] - 1, :, :]) / tv_y_size) +
(tf.nn.l2_loss(image[:, :, 1:, :] - image[:, :, :content_shape[2] - 1, :]) / tv_x_size))
# overall loss
loss = content_loss + style_loss_for_all_styles + tv_loss
# optimizer
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
def print_progress(i, last=False):
stderr.write('Iteration %d/%d' % (i + 1, iterations))
if last or (print_iterations and i % print_iterations == 0):
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss_for_all_styles.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf') # all losses will be lower than initial
best = None
total_initial_time = datetime.now().replace(microsecond=0)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
initial_time = datetime.now().replace(microsecond=0)
for i in range(iterations):
now_time = datetime.now().replace(microsecond=0)
last_step = (i == iterations - 1)
print_progress(i, last=last_step)
stderr.write(' Training Time %s Elapsed Time %s\n' % (str(now_time - initial_time), str(now_time - total_initial_time)))
initial_time = now_time
train_step.run()
# when checkpoint_iterations is not None
# and when iter idx fulfills it
# or when it comes to the last step
if(checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
# image was a tf.Variable
# eval it turns it into numbers, I guess
best = image.eval()
# what is yield?
# content_shape[1:] means
# shape[1], shape[2], shape[3]
# but eliminate the shape[0] which
# means the number of images
yield(
(None if last_step else i),
(vgg.unprocess(best.reshape(content_shape[1:]), mean_pixel)))
def train(content_targets,
style_target,
content_weight,
style_weight,
tv_weight,
vgg_path,
epochs=2,
print_iterations=1000,
batch_size=4,
learning_rate=1e-3,
save_path='model/style.ckpt'):
# 根据batch丢弃最后的训练图像
mod = len(content_targets) % batch_size
if mod > 0:
content_targets = content_targets[:-mod]
style_features = {}
# 训练图像大小：320x320x3，按照tensorflow格式
batch_shape = (batch_size, 320, 320, 3)
style_shape = (1, ) + style_target.shape
# 读取训练好的VGGNet模型
weights, mean_pixel = vgg.load_net(vgg_path)
with tf.Graph().as_default(), tf.Session() as sess:
style_image = tf.placeholder(tf.float32,
shape=style_shape,
name='style_image')
# 没看错！空图片减去均值
style_image_pre = vgg.preprocess(style_image, mean_pixel)
net = vgg.net(weights, style_image_pre)
# 把style图片展开形成数组
style_pre = np.array([style_target])
for layer in STYLE_LAYER:
# 取出该层的计算结果
features = net[layer].eval(feed_dict={style_image: style_pre})
# 行数为该层的Filter数（参见论文）
features = np.reshape(features, (-1, features.shape[3]))
# Gram Matrix: A'A (参见论文)
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
with tf.Graph().as_default(), tf.Session() as sess:
x_content = tf.placeholder(tf.float32,
shape=batch_shape,
name='x_content')
x_pre = vgg.preprocess(x_content, mean_pixel)
content_features = {}
content_net = vgg.net(weights, x_pre)
# 同上，提取所需层
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER]
# 使用残差网络
preds = residual.net(x_content / 255.0)
preds_pre = vgg.preprocess(preds, mean_pixel)
net = vgg.net(weights, preds_pre)
# 计算每个batch里的所有数据
content_size = _tensor_size(
content_features[CONTENT_LAYER]) * batch_size
assert _tensor_size(content_features[CONTENT_LAYER]) == _tensor_size(
net[CONTENT_LAYER])
# 计算经过残差网络和不经过时的差别
content_loss = content_weight * (
2 * tf.nn.l2_loss(net[CONTENT_LAYER] -
content_features[CONTENT_LAYER]) / content_size)
# 计算经过残差网络的图像与style图像之间的差别
style_losses = []
for style_layer in STYLE_LAYER:
layer = net[style_layer]
bs, height, width, filters = map(lambda i: i.value,
layer.get_shape())
size = height * width * filters
feats = tf.reshape(layer, (bs, height * width, filters))
feats_T = tf.transpose(feats, perm=[0, 2, 1])
# Gram Matrix: A'A (参见论文)
grams = tf.matmul(feats_T, feats) / size
style_gram = style_features[style_layer]
style_losses.append(2 * tf.nn.l2_loss(grams - style_gram) /
style_gram.size)
style_loss = style_weight * functools.reduce(tf.add,
style_losses) / batch_size
# 去图像噪声: Total Variation
tv_y_size = _tensor_size(preds[:, 1:, :, :])
tv_x_size = _tensor_size(preds[:, :, 1:, :])
y_tv = tf.nn.l2_loss(preds[:, 1:, :, :] -
preds[:, :batch_shape[1] - 1, :, :])
x_tv = tf.nn.l2_loss(preds[:, :, 1:, :] -
preds[:, :, :batch_shape[2] - 1, :])
tv_loss = tv_weight * 2 * (x_tv / tv_x_size +
y_tv / tv_y_size) / batch_size
# 最终的loss函数
loss = content_loss + style_loss + tv_loss
# 开始训练过程
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
sess.run(tf.global_variables_initializer())
for epoch in range(epochs):
num_examples = len(content_targets)
iterations = 0
start_time = time.time()
# 每一次epoch就用训练集的所有图片训练一遍
while iterations * batch_size < num_examples:
curr = iterations * batch_size
step = curr + batch_size
X_batch = np.zeros(batch_shape, dtype=np.float32)
for j, img_p in enumerate(content_targets[curr:step]):
X_batch[j] = helper.read_img(img_p, (320, 320, 3)).astype(
np.float32)
iterations += 1
# 确保每批次计算的时候不出错
assert X_batch.shape[0] == batch_size
feed_dict = {x_content: X_batch}
# 开始训练
train_step.run(feed_dict=feed_dict)
# 隔几次打印一次训练进度
is_print_iter = int(iterations) % print_iterations == 0
# 是否是最后一个epoch
is_last = epoch == epochs - 1 and iterations * batch_size >= num_examples
# 打印信息
should_print = is_print_iter or is_last
if should_print:
current_time = time.time()
delta_time = current_time - start_time
start_time = current_time
to_get = [style_loss, content_loss, tv_loss, loss, preds]
test_feed_dict = {x_content: X_batch}
tup = sess.run(to_get, feed_dict=test_feed_dict)
_style_loss, _content_loss, _tv_loss, _loss, _preds = tup
losses = (_style_loss, _content_loss, _tv_loss, _loss)
saver = tf.train.Saver()
res = saver.save(sess, save_path)
yield (_preds, losses, iterations, epoch, delta_time)
def optimize(content_targets, style_target, content_weight, style_weight,
tv_weight, vgg_path, epochs=2, print_iterations=1000, batch_size=4,
checkpoint_dir='saver/fns.ckpt', summary_dir='summary/', learning_rate=1e-3):
"""
Calculate the total loss and optimize the network.
Args:
content_targets: The content image.
style_target: The style image.
content_weight: Weight for content loss.
style_weight: Weight for style loss.
tv_weight: Weight for total vaiance.
vgg_path: Path of the vgg network.
epochs: Number of epochs for training. Default: 2.
print_iteration: Print the trainging loss. Default: 1000
batch_size: Default: 4.
checkpoint_dir: Path to save the checkpoint.
summary_dir: Path to save summaries.
learning_rate: Default: 1e-3.
Returns:
Yield the prediction, losses, iteration and epoch
"""
mod = len(content_targets) % batch_size
if mod > 0:
print("Train set has been trimmed slightly..")
content_targets = content_targets[:-mod] # discard the remaining
batch_shape = (batch_size, 256, 256, 3)
# precompute style features
style_features = _style_features(style_target, vgg_path)
X_content = tf.placeholder(tf.float32, shape=batch_shape, name="X_content")
X_pre = vgg.preprocess(X_content)
# compute content features
content_features = {}
content_net = vgg.net(vgg_path, X_pre)
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER]
# content is the input for both the transform network and the loss
# network
preds = transform.net(X_content/255.0)
preds_pre = vgg.preprocess(preds)
net = vgg.net(vgg_path, preds_pre)
# compute loss
content_loss = _content_loss(content_weight, net, content_features, batch_size)
style_loss = _style_loss(style_weight, net, style_features)
tv_loss = _tv_loss(tv_weight, preds, batch_shape)
loss = content_loss + style_loss + tv_loss
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
# summary for tensorboard
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_op = tf.summary.merge_all()
writer = tf.summary.FileWriter(summary_dir, graph=tf.get_default_graph())
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(epochs):
num_examples = len(content_targets)
iterations = 0
while iterations * batch_size < num_examples:
curr = iterations * batch_size
step = curr + batch_size
X_batch = np.zeros(batch_shape, dtype=np.float32)
for j, img_p in enumerate(content_targets[curr:step]):
X_batch[j] = get_img(img_p, (256, 256, 3)).astype(np.float32) #resize content image
iterations += 1
assert X_batch.shape[0] == batch_size
feed_dict = {
X_content : X_batch
}
#train_step.run(feed_dict = feed_dict)
summary, _ = sess.run([summary_op, train_step], feed_dict = feed_dict)
is_print_iter = int(iterations) % print_iterations == 0
is_last = epoch == epochs - 1 and iterations * batch_size >= num_examples
should_print = is_print_iter or is_last
if should_print:
to_get = [style_loss, content_loss, tv_loss, loss, preds]
test_feed_dict = {X_content:X_batch}
tup = sess.run(to_get, feed_dict = test_feed_dict)
style_loss_p, content_loss_p, tv_loss_p,loss_p, preds_p = tup
losses = (style_loss_p, content_loss_p, tv_loss_p, loss_p)
saver = tf.train.Saver(max_to_keep = 5)
res = saver.save(sess, checkpoint_dir, iterations)
yield(preds_p, losses, iterations, epoch)
if int(iterations) % 20 == 0:
writer.add_summary(summary)
def stylize(network, initial, initial_noiseblend, content, styles, preserve_colors, iterations,
content_weight, content_weight_blend, style_weight, style_layer_weight_exp, style_blend_weights, tv_weight,
learning_rate, beta1, beta2, epsilon, pooling,
print_iterations=None, checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, pooling)
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features_bank = sk_image.extract_patches_2d(np.squeeze(features), (kernel_s, kernel_s))
style_features[i][layer] = [features_bank,features]
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
# noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (tf.random_normal(shape) * 0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_layers_weights = {}
# content_layers_weights['relu4_2'] = content_weight_blend
# content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_layers_weights['relu4_2'] = 0.5
content_layers_weights['relu5_2'] = 0.5
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(content_layers_weights[content_layer] * content_weight * (2 * tf.nn.l2_loss(
net[content_layer] - content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
# Calculate normalized layer
layer = tf.expand_dims(net[style_layer],axis=4)
paddings = [[0, 0], [1,1], [1,1], [0, 0],[0,0]]
layer_depth = layer.get_shape().as_list()[3]
layer_pad = tf.pad(layer, paddings, "CONSTANT")
layer_norm = tf.sqrt(tf.nn.conv3d(tf.pow(layer_pad,2),tf.ones((kernel_s,kernel_s,layer_depth,1,1),dtype=tf.float32),strides=[1, 1, 1, 1, 1],padding='VALID'))
# Calculate normalized filter bank
style_filters = np.transpose(style_features[i][style_layer][0],(1,2,3,0))
style_filters = np.expand_dims(style_filters,axis=3)
style_filters_norm = np.sqrt(np.sum(np.power(style_filters,2),axis=(0,1,2)))
style_filters_normalized = style_filters/style_filters_norm
# Calculate normalized correlations
layer_filtered = tf.nn.conv3d(layer_pad,style_filters_normalized,strides=[1, 1, 1, 1, 1],padding='VALID')/layer_norm
# Find maximum response and index into the filters
max_filter_response_idx = tf.squeeze(tf.argmax(layer_filtered,axis=4))
# max_filter_response_idx = tf.squeeze(tf.argmax(tf.abs(layer_filtered),axis=4))
max_filter_response_idx = tf.reshape(max_filter_response_idx,[-1])
max_filter_response_weight = tf.squeeze(tf.reduce_max(tf.abs(layer_filtered),axis=4))
max_filter_response_weight = tf.reshape(max_filter_response_weight,[-1])
max_filter_response_weight = max_filter_response_weight/tf.reduce_max(max_filter_response_weight)
style_filters_tf = tf.transpose(tf.squeeze(tf.convert_to_tensor(style_filters, np.float32)),(3,0,1,2))
style_filters_tf_gathered = tf.gather(style_filters_tf,max_filter_response_idx)
style_filters_tf_gathered = tf.reshape(style_filters_tf_gathered,(style_filters_tf_gathered.get_shape().as_list()[0], -1))
layer_patches = tf.extract_image_patches(tf.squeeze(layer_pad,axis=4),
[1,kernel_s,kernel_s,1],
[1,1,1,1],
[1,1,1,1],
padding="VALID")
layer_size = tf.shape(layer_patches)
layer_patches = tf.reshape(layer_patches,(-1, layer_size[3]))
style_norm = tf.cast(layer_size[1]*layer_size[2]*layer_size[3],dtype=tf.float32)
# gram1 = tf.matmul(tf.transpose(layer_patches), layer_patches) / style_norm
# gram2 = tf.matmul(tf.transpose(style_filters_tf_gathered), style_filters_tf_gathered) / style_norm
# style_losses.append(style_layers_weights[style_layer] * 2 * tf.nn.l2_loss(gram1- gram2))
loss_ = tf.reduce_mean(tf.reduce_mean(tf.pow(layer_patches-style_filters_tf_gathered, 2),axis=1)*tf.stop_gradient(max_filter_response_weight))
style_losses.append(style_layers_weights[style_layer] * 2 * loss_)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2, epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in range(iterations):
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
# print(str(max_filter_response_weight.eval()))
# print(' ')
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations and i % print_iterations == 0):
print_progress()
if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(Image.fromarray(original_image.astype(np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(Image.fromarray(combined_yuv, 'YCbCr').convert('RGB'))
yield (
(None if last_step else i),
img_out
)
G_sample) # G_sample = generated dataset (fake)
# D_real & D_fake = unused (D_fake = probability G fools D)
""" Feature Loss """
#VGG
#content = imread('abbeyexample_copy.png')/256
#content = gray2rgb(rgb2gray(content))
shape = (1, 256, 256, 3)
pooling = 'avg'
CONTENT_LAYERS = ('relu4_2', 'relu5_2')
network = 'imagenet-vgg-verydeep-19.mat'
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
print(5)
orig_image = tf.placeholder(
'float', shape=shape) #need to feed it with (1,256,256,3) objects
print(orig_image)
orig_content = vgg.preprocess(orig_image,
vgg_mean_pixel) #tensor (1,256,256,3)
print(orig_content)
print('G_sample.shape', G_sample.shape)
G_sample_dim = G_sample
G_sample = tf.reshape(G_sample, (256, 256))
G_sample = tf.stack([G_sample, G_sample, G_sample],
axis=2) #tensor (256,256,3)
print('G_sample.shape', G_sample.shape)
gen_content = vgg.preprocess(G_sample, vgg_mean_pixel)
gen_content = tf.expand_dims(gen_content, 0)
print('ok')
orig_net = vgg.net_preloaded(vgg_weights, orig_content, pooling)
print('ok1')
gen_net = vgg.net_preloaded(vgg_weights, gen_content, pooling)
#content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
#print('content_pre.shape',content_pre.shape)
def optimize(content_targets, style_target, content_weight, style_weight,
tv_weight, vgg_path, epochs=2, print_iterations=1000,
batch_size=4, save_path='saver/fns.ckpt', slow=False,
learning_rate=1e-3, debug=False):
if slow:
batch_size = 1
mod = len(content_targets) % batch_size
if mod > 0:
print("Train set has been trimmed slightly..")
content_targets = content_targets[:-mod]
style_features = {}
batch_shape = (batch_size,256,256,3)
style_shape = (1,) + style_target.shape
print(style_shape)
# precompute style features
with tf.Graph().as_default(), tf.device('/cpu:0'), tf.Session() as sess:
style_image = tf.placeholder(tf.float32, shape=style_shape, name='style_image')
style_image_pre = vgg.preprocess(style_image)
net = vgg.net(vgg_path, style_image_pre)
style_pre = np.array([style_target])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={style_image:style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
with tf.Graph().as_default(), tf.Session() as sess:
X_content = tf.placeholder(tf.float32, shape=batch_shape, name="X_content")
X_pre = vgg.preprocess(X_content)
# precompute content features
content_features = {}
content_net = vgg.net(vgg_path, X_pre)
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER]
if slow:
preds = tf.Variable(
tf.random_normal(X_content.get_shape()) * 0.256
)
preds_pre = preds
else:
preds = transform.net(X_content/255.0)
preds_pre = vgg.preprocess(preds)
net = vgg.net(vgg_path, preds_pre)
content_size = _tensor_size(content_features[CONTENT_LAYER])*batch_size
assert _tensor_size(content_features[CONTENT_LAYER]) == _tensor_size(net[CONTENT_LAYER])
content_loss = content_weight * (2 * tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) / content_size
)
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
bs, height, width, filters = map(lambda i:i.value,layer.get_shape())
size = height * width * filters
feats = tf.reshape(layer, (bs, height * width, filters))
feats_T = tf.transpose(feats, perm=[0,2,1])
grams = tf.matmul(feats_T, feats) / size
style_gram = style_features[style_layer]
style_losses.append(2 * tf.nn.l2_loss(grams - style_gram)/style_gram.size)
style_loss = style_weight * functools.reduce(tf.add, style_losses) / batch_size
# total variation denoising
tv_y_size = _tensor_size(preds[:,1:,:,:])
tv_x_size = _tensor_size(preds[:,:,1:,:])
y_tv = tf.nn.l2_loss(preds[:,1:,:,:] - preds[:,:batch_shape[1]-1,:,:])
x_tv = tf.nn.l2_loss(preds[:,:,1:,:] - preds[:,:,:batch_shape[2]-1,:])
tv_loss = tv_weight*2*(x_tv/tv_x_size + y_tv/tv_y_size)/batch_size
loss = content_loss + style_loss + tv_loss
# overall loss
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
sess.run(tf.global_variables_initializer())
import random
uid = random.randint(1, 100)
print("UID: %s" % uid)
for epoch in range(epochs):
num_examples = len(content_targets)
iterations = 0
while iterations * batch_size < num_examples:
start_time = time.time()
curr = iterations * batch_size
step = curr + batch_size
X_batch = np.zeros(batch_shape, dtype=np.float32)
for j, img_p in enumerate(content_targets[curr:step]):
X_batch[j] = get_img(img_p, (256,256,3)).astype(np.float32)
iterations += 1
assert X_batch.shape[0] == batch_size
feed_dict = {
X_content:X_batch
}
train_step.run(feed_dict=feed_dict)
end_time = time.time()
delta_time = end_time - start_time
if debug:
print("UID: %s, batch time: %s" % (uid, delta_time))
is_print_iter = int(iterations) % print_iterations == 0
if slow:
is_print_iter = epoch % print_iterations == 0
is_last = epoch == epochs - 1 and iterations * batch_size >= num_examples
should_print = is_print_iter or is_last
if should_print:
to_get = [style_loss, content_loss, tv_loss, loss, preds]
test_feed_dict = {
X_content:X_batch
}
tup = sess.run(to_get, feed_dict = test_feed_dict)
_style_loss,_content_loss,_tv_loss,_loss,_preds = tup
losses = (_style_loss, _content_loss, _tv_loss, _loss)
if slow:
_preds = vgg.unprocess(_preds)
else:
saver = tf.train.Saver()
res = saver.save(sess, save_path)
yield(_preds, losses, iterations, epoch)
def stylize(network, initial, initial_noiseblend, content, styles, preserve_colors, iterations,
content_weight, content_weight_blend, style_weight, style_layer_weight_exp, style_blend_weights, tv_weight,
learning_rate, beta1, beta2, epsilon, pooling,
print_iterations=None, checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1,) + content.shape #若content.shape=(356, 600, 3) shape=(1，356, 600, 3)
style_shapes = [(1,) + style.shape for style in styles]
content_features = {} #创建内容features map
style_features = [{} for _ in styles] #创建风格features map
vgg_weights, vgg_mean_pixel = vgg.load_net(network) #加载预训练模型，得到weights和mean_pixel
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp #若有设置style_layer_weight_exp，则style_layers_weights指数增长，
# style_layer_weight_exp默认为1不增长
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum #更新style_layers_weights应该是比例，使其总和为1
# 首先创建一个image的占位符，然后通过eval()的feed_dict将content_pre传给image，
# 启动net的运算过程，得到了content的feature maps
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.compat.v1.Session() as sess: #计算content features
image = tf.compat.v1.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling) #所有网络在此构建，net为content的features maps
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)]) #content - vgg_mean_pixel
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(feed_dict={image: content_pre}) #content_features取值
# print(layer,content_features[layer].shape)
# compute style features in feedforward mode
for i in range(len(styles)): #计算style features
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.compat.v1.Session() as sess:
image = tf.compat.v1.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, pooling) #pooling 默认为MAX
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)]) #styles[i]-vgg_mean_pixel
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3])) #根据通道数目reshape
gram = np.matmul(features.T, features) / features.size #gram矩阵
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256 #初始化图片
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + ( tf.random.normal(shape) * 0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
'''
image = tf.Variable(initial)初始化了一个TensorFlow的变量，即为我们需要训练的对象。
注意这里我们训练的对象是一张图像，而不是weight和bias。
'''
net = vgg.net_preloaded(vgg_weights, image, pooling) #此处的net为生成图片的features map
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend #内容图片 content weight blend, conv4_2 * blend + conv5_2 * (1-blend)
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend #content weight blend默认为1，即只用conv4_2层
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(content_layers_weights[content_layer] * content_weight * (2 * tf.nn.l2_loss(
net[content_layer] - content_features[content_layer]) / #生成图片-内容图片
content_features[content_layer].size)) # tf.nn.l2_loss：output = sum(t ** 2) / 2
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
'''
由于style图像可以输入多幅，这里使用for循环。同样的，将style_pre传给image占位符，
启动net运算，得到了style的feature maps，由于style为不同filter响应的内积，
因此在这里增加了一步：gram = np.matmul(features.T, features) / features.size，即为style的feature。
'''
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size #求得生成图片的gram矩阵
style_gram = style_features[i][style_layer]
style_losses.append(style_layers_weights[style_layer] * 2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
'''
接下来定义了Content Loss和Style Loss，结合文中的公式很容易看懂，在代码中，
还增加了total variation denoising，因此总的loss = content_loss + style_loss + tv_loss
'''
loss = content_loss + style_loss + tv_loss #总loss为三个loss之和
# optimizer setup
# optimizer setup
# 创建train_step，使用Adam优化器，优化对象是上面的loss
# 优化过程，通过迭代使用train_step来最小化loss，最终得到一个best，即为训练优化的结果
train_step = tf.compat.v1.train.AdamOptimizer(learning_rate, beta1, beta2, epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.compat.v1.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in range(iterations):
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations and i % print_iterations == 0):
print_progress()
if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel) #还原图片
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(Image.fromarray(original_image.astype(np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(Image.fromarray(combined_yuv, 'YCbCr').convert('RGB'))
yield ( #相当于return，但用于迭代
(None if last_step else i),
img_out
)
def main():
content_path, style_path, width, style_scale = sys.argv[1:]
width = int(width)
style_scale = float(style_scale)
content_image = imread(content_path)
style_image = imread(style_path)
if width > 0:
new_shape = (int(math.floor(float(content_image.shape[0]) /
content_image.shape[1] * width)), width)
content_image = sm.imresize(content_image, new_shape)
if style_scale > 0:
style_image = sm.imresize(style_image, style_scale)
shape = (1,) + content_image.shape
style_shape = (1,) + style_image.shape
content_features = {}
style_features = {}
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net, mean_pixel = vgg.net(VGG_PATH, image)
content_pre = np.array([vgg.preprocess(content_image, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(
feed_dict={image: content_pre})
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shape)
net, _ = vgg.net(VGG_PATH, image)
style_pre = np.array([vgg.preprocess(style_image, mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
grammatrix = np.matmul(features.T, features)
style_features[layer] = grammatrix
g = tf.Graph()
with g.as_default():
global_step = tf.Variable(0, trainable=False)
noise = np.random.normal(size=shape, scale=np.std(content_image) * 0.1)
content_pre = vgg.preprocess(content_image, mean_pixel)
init = content_pre * (1 - NOISE_RATIO) + noise * NOISE_RATIO
init = init.astype('float32')
image = tf.Variable(init)
net, _ = vgg.net(VGG_PATH, image)
content_loss = tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER])
style_losses = []
for i in STYLE_LAYERS:
layer = net[i]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats)
style_gram = style_features[i]
style_losses.append(tf.nn.l2_loss(gram - style_gram) /
(4.0 * number ** 2 * (height * width) ** 2))
style_loss = reduce(tf.add, style_losses) / len(style_losses)
loss = ALPHA * content_loss + BETA * style_loss
learning_rate = tf.train.exponential_decay(LEARNING_RATE_INITIAL,
global_step, LEARNING_DECAY_STEPS, LEARNING_DECAY_BASE,
staircase=True)
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss,
global_step=global_step)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(100000):
print 'i = %d' % i
imsave('%05d.jpg' % i, vgg.unprocess(
image.eval().reshape(shape[1:]), mean_pixel))
train_step.run()
VGG_PATH) # convert the tf.nn.conv2d to slim format of vgg
decoder = Decoder(mode='test', weights_path=DECODER_PATH)
content_input = tf.placeholder(tf.float32,
shape=(1, None, None, 3),
name='content_input')
style_input = tf.placeholder(tf.float32,
shape=(1, None, None, 3),
name='style_input')
# switch RGB to BGR
content = tf.reverse(content_input, axis=[-1])
style = tf.reverse(style_input, axis=[-1])
# preprocess image
content = vgg.preprocess(content)
style = vgg.preprocess(style)
encoder_content, encoder_content_points = vgg.vgg_19(
content, reuse=False, final_endpoint="conv4_1")
encoder_style, encoder_style_points = vgg.vgg_19(
style, reuse=True, final_endpoint="conv4_1")
# pass the encoded images to AdaIN
target_features = AdaIN(encoder_content, encoder_style)
# decode target features back to image
with tf.variable_scope("decoder_target"):
#alpha = 0.8
#target_features=(1-alpha)*encoder_content+alpha*target_features #content-style trade-off
generated_img = decoder.decode(target_features)
def stylize(network, initial, initial_noiseblend, content, styles, preserve_colors, iterations,
content_weight, content_weight_blend, style_weight, style_layer_weight_exp, style_blend_weights, tv_weight,
learning_rate, beta1, beta2, epsilon, pooling,
print_iterations=None, checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image);
`iteration` is None if this is the final image (the last iteration).
Otherwise tuples are yielded every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
# The shape information in the comment is based on the content image 1-content.jpg with shape (533, 400, 3)
# and 1-style.jpg (316, 400, 3)
# This should be changed with different images.
shape = (1,) + content.shape # (1, 533, 400, 3)
style_shapes = [(1,) + style.shape for style in styles] # (1, 316, 400, 3)
content_features = {}
style_features = [{} for _ in styles]
vgg_weights, vgg_mean_pixel = vgg.load_net(network) # Load the VGG-19 model.
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight # {'relu1_1': 1.0, 'relu2_1': 1.0, 'relu3_1': 1.0, 'relu4_1': 1.0, 'relu5_1': 1.0}
layer_weight *= style_layer_weight_exp # 1.0
# VGG19 layers:
# 'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1',
# 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',
# 'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'conv3_4', 'relu3_4', 'pool3',
# 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3', 'relu4_3', 'conv4_4', 'relu4_4', 'pool4',
# 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3', 'conv5_4', 'relu5_4'
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS: # ('relu1_1', 'relu2_1', 'relu3_1', 'relu4_1', 'relu5_1')
layer_weights_sum += style_layers_weights[style_layer] # 5.0
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum # {'relu1_1': 0.2, 'relu2_1': 0.2, 'relu3_1': 0.2, 'relu4_1': 0.2, 'relu5_1': 0.2}
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling) # {'conv1_1': Tensor..., relu1_1: Tensor...}
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)]) # (1, 533, 400, 3) subtract with the mean pixel
for layer in CONTENT_LAYERS: # (relu4_2, relu5_2)
content_features[layer] = net[layer].eval(feed_dict={image: content_pre}) # Find the feature values for (relu4_2, relu5_2)
# compute style features in feed forward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i]) # (1, 316, 400, 3)
net = vgg.net_preloaded(vgg_weights, image, pooling)
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS: # # ('relu1_1', 'relu2_1', 'relu3_1', 'relu4_1', 'relu5_1')
features = net[layer].eval(feed_dict={image: style_pre}) # For relu1_1 layer (1, 316, 400, 64)
features = np.reshape(features, (-1, features.shape[3])) # (126400, 64)
gram = np.matmul(features.T, features) / features.size # (64, 64) Gram matrix - measure the dependency of features.
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend # 0
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1) # Generate a random image with SD the same as the content image.
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (tf.random_normal(shape) * 0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS: # {'relu5_2'}
# Use MSE as content losses
content_losses.append(content_layers_weights[content_layer] * content_weight * (2 * tf.nn.l2_loss(
net[content_layer] - content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer] # For relu1_1: (1, 533, 400, 64)
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number)) # (213200, 64)
gram = tf.matmul(tf.transpose(feats), feats) / size # Gram matrix for the features in relu1_1 for the result image.
style_gram = style_features[i][style_layer] # Gram matrix for the style
# Style loss is the MSE for the difference of the 2 Gram matrix
style_losses.append(style_layers_weights[style_layer]
* 2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# Total variation denoising: Add cost to penalize neighboring pixel is very different.
# This help to reduce noise.
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2, epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in range(iterations):
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations and i % print_iterations == 0):
print_progress()
if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(Image.fromarray(original_image.astype(np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(Image.fromarray(combined_yuv, 'YCbCr').convert('RGB'))
yield (
(None if last_step else i),
img_out
)
def stylize(network, initial, content, styles, iterations,
content_weight, style_weight, style_blend_weights, tv_weight,
learning_rate, print_iterations=None, checkpoint_iterations=None):
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net, mean_pixel = vgg.net(network, image)
content_pre = np.array([vgg.preprocess(content, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net, _ = vgg.net(network, image)
style_pre = np.array([vgg.preprocess(styles[i], mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
print 'Initial feature shape: ', features.shape
features = np.reshape(features, (-1, features.shape[3]))
#mask = np.zeros_like(features)
#mask[:49664/2, :] = 1
#print 'Mask shape', mask.shape
print 'Final features shape', features.shape
#features = features*mask
gram = np.matmul(features.T, features) / features.size
print 'Gram matrix shape: ', gram.shape
style_features[i][layer] = gram
#sys.exit()
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, mean_pixel)])
initial = initial.astype('float32')
image = tf.Variable(initial)
net, _ = vgg.net(network, image)
# content loss
content_loss = content_weight * (2 * tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) /
content_features[CONTENT_LAYER].size)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
print 'Height, width, number', height, width, number
size = height * width * number
feats = tf.reshape(layer, (-1, number))
#print tf.shape(feats).as_list()
if normal_flag == 0:
mask = np.zeros((height*width, number), dtype=np.float32)
maskt = np.reshape(imread('bottle_mask.jpg').astype(np.float32), (height*width,))
maskt = maskt > 100
for d in xrange(number):
mask[:,d] = maskt
print 'Mask shape', mask.shape
#print sum(sum(mask == 1)) + sum(sum(mask == 0))
#mask[:height*width/2, :] = 1
if i == 0:
mask = tf.constant(mask)
feats = tf.mul(feats,mask)
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
else:
mask2 = mask < 1
feats2 = tf.mul(feats,mask2)
gram2 = tf.matmul(tf.transpose(feats2), feats2) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram2 - style_gram) / style_gram.size)
else:
feats2 = feats
gram2 = tf.matmul(tf.transpose(feats2), feats2) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram2 - style_gram) / style_gram.size)
pass
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
if normal_flag != 0:
print "general mask :"
mask = np.zeros((height*width, number), dtype=np.float32)
maskt = np.reshape(imread('bottle_mask.jpg').astype(np.float32), (height*width,))
maskt = maskt > 100
# for d in xrange(3):
# mask[:,d] = maskt
print 'Mask shape', maskt.shape
maskt = maskt.reshape((height,width))
maskt = np.array([maskt,maskt,maskt])
maskt = maskt.transpose((1,2,0))
mask = tf.constant(maskt, dtype=tf.float32)
# feats = tf.mul(feats,mask)
def capper(a,b,mask):
# (1, 468, 304, 3)
print "orig shape", a
reshaped_in_grad = tf.reshape(a,[-1] )
print "reshaped grad", reshaped_in_grad
print "mask" ,mask
g = tf.mul(a,mask)
# g = tf.reshape(g, (1,height,width,3))
# print a,b
# print g
return g,b
# optimizer setup
# train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
# # Create an optimizer.
train_step = tf.train.GradientDescentOptimizer(learning_rate)
# # Compute the gradients for a list of variables.
grads_and_vars = train_step.compute_gradients(loss)
# # grads_and_vars is a list of tuples (gradient, variable). Do whatever you
# # need to the 'gradient' part, for example cap them, etc.
capped_grads_and_vars = [(capper(gv[0], gv[1], mask)) for gv in grads_and_vars]
# # Ask the optimizer to apply the capped gradients.
train_step = train_step.apply_gradients(capped_grads_and_vars)
# opt_op = opt.minimize(cost, var_list=<list of variables>)
def print_progress(i, last=False):
if print_iterations is not None:
if i is not None and i % print_iterations == 0 or last:
print >> stderr, ' content loss: %g' % content_loss.eval()
print >> stderr, ' style loss: %g' % style_loss.eval()
print >> stderr, ' tv loss: %g' % tv_loss.eval()
print >> stderr, ' total loss: %g' % loss.eval()
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(iterations):
print_progress(i)
print >> stderr, 'Iteration %d/%d' % (i + 1, iterations)
train_step.run()
# print "runningstep: ",i, running_step
if (checkpoint_iterations is not None and
i % checkpoint_iterations == 0) or i == iterations - 1:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
print_progress(None, i == iterations - 1)
if i % 10 == 0 and best is not None:
tmp_img = vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
imsave("iter" + str(i) + ".jpg", tmp_img)
return vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
def stylize(network,
initial,
initial_noiseblend,
content,
styles,
preserve_colors,
iterations,
content_weight,
content_weight_blend,
style_weight,
style_layer_weight_exp,
style_blend_weights,
tv_weight,
learning_rate,
beta1,
beta2,
epsilon,
pooling,
print_iterations=None,
checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image, loss_vals) at every
iteration. However `image` and `loss_vals` are None by default. Each
`checkpoint_iterations`, `image` is not None. Each `print_iterations`,
`loss_vals` is not None.
`loss_vals` is a dict with loss values for the current iteration, e.g.
``{'content': 1.23, 'style': 4.56, 'tv': 7.89, 'total': 13.68}``.
:rtype: iterator[tuple[int,image]]
"""
shape = (1, ) + content.shape
style_shapes = [(1, ) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, pooling)
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (
tf.random_normal(shape) *
0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(
content_layers_weights[content_layer] * content_weight *
(2 * tf.nn.l2_loss(net[content_layer] -
content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value,
layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(style_layers_weights[style_layer] * 2 *
tf.nn.l2_loss(gram - style_gram) /
style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(
tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:, 1:, :, :])
tv_x_size = _tensor_size(image[:, :, 1:, :])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:, 1:, :, :] - image[:, :shape[1] - 1, :, :])
/ tv_y_size) +
(tf.nn.l2_loss(image[:, :, 1:, :] - image[:, :, :shape[2] - 1, :])
/ tv_x_size))
# total loss
loss = content_loss + style_loss + tv_loss
# We use OrderedDict to make sure we have the same order of loss types
# (content, tv, style, total) as defined by the initial costruction of
# the loss_store dict. This is important for print_progress() and
# saving loss_arrs (column order) in the main script.
#
# Subtle Gotcha (tested with Python 3.5): The syntax
# OrderedDict(key1=val1, key2=val2, ...) does /not/ create the same
# order since, apparently, it first creates a normal dict with random
# order (< Python 3.7) and then wraps that in an OrderedDict. We have
# to pass in a data structure which is already ordered. I'd call this a
# bug, since both constructor syntax variants result in different
# objects. In 3.6, the order is preserved in dict() in CPython, in 3.7
# they finally made it part of the language spec. Thank you!
loss_store = OrderedDict([('content', content_loss),
('style', style_loss), ('tv', tv_loss),
('total', loss)])
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print('Optimization started...')
if (print_iterations and print_iterations != 0):
print_progress(get_loss_vals(loss_store))
iteration_times = []
start = time.time()
for i in range(iterations):
iteration_start = time.time()
if i > 0:
elapsed = time.time() - start
# take average of last couple steps to get time per iteration
remaining = np.mean(
iteration_times[-10:]) * (iterations - i)
print('Iteration %4d/%4d (%s elapsed, %s remaining)' %
(i + 1, iterations, hms(elapsed), hms(remaining)))
else:
print('Iteration %4d/%4d' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations
and i % print_iterations == 0):
loss_vals = get_loss_vals(loss_store)
print_progress(loss_vals)
else:
loss_vals = None
if (checkpoint_iterations
and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]),
vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(
Image.fromarray(
styled_grayscale_rgb.astype(
np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(
Image.fromarray(original_image.astype(
np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(
Image.fromarray(combined_yuv,
'YCbCr').convert('RGB'))
if iterations % 2 == 0:
print(iterations)
print('dang o day')
Image.fromarray(img).save('output' +
str(iterations) + '.jpg',
quality=95)
else:
img_out = None
yield i + 1 if last_step else i, img_out, loss_vals
iteration_end = time.time()
iteration_times.append(iteration_end - iteration_start)
def inferenceImg(network, initial_img, initial_noiseblend, content, style,
preserve_colors, iterations, content_weight,
content_weight_blend, style_weight, style_layer_weight_exp,
style_blend_weight, tv_weight, learning_rate, beta1, beta2,
epsilon, pooling, print_iterations, checkpoint_iterations):
content_shape = (1, ) + content.shape
style_shape = (1, ) + style.shape
content_features = {}
style_features = {}
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight = layer_weight * style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum = layer_weights_sum + style_layers_weights[
style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = style_layers_weights[
style_layer] / layer_weights_sum
# compute content features in feedforward mode
g1 = tf.Graph()
with g1.as_default(), g1.device('/cpu:0'), tf.Session() as sess:
contentImg = tf.placeholder('float', shape=content_shape)
net = vgg.net_preloaded(vgg_weights, contentImg, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(
feed_dict={contentImg: content_pre})
# compute style features in feedforward mode
g2 = tf.Graph()
with g2.as_default(), g2.device('/cpu:0'), tf.Session() as sess:
styleImg = tf.placeholder('float', shape=style_shape)
net = vgg.net_preloaded(vgg_weights, styleImg, pooling)
style_pre = np.array([vgg.preprocess(style, vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={styleImg: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
noise = np.random.normal(size=content_shape,
scale=np.std(content) * 0.1)
initial = tf.random_normal(content_shape) * 0.256
inferenceImg = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, inferenceImg, pooling)
# compute content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(
content_layers_weights[content_layer] * content_weight *
(2 * tf.nn.l2_loss(net[content_layer] -
content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# compute style loss
style_loss = 0
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value,
layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[style_layer]
style_losses.append(style_layers_weights[style_layer] * 2 *
tf.nn.l2_loss(gram - style_gram) /
style_gram.size)
style_loss += style_weight * style_blend_weight * reduce(
tf.add, style_losses)
# skip compute variation denoise, in order to shorten the running time
# total variation denoising
# tv_y_size = _tensor_size(inferenceImg[:, 1:, :, :])
# tv_x_size = _tensor_size(inferenceImg[:, :, 1:, :])
# tv_loss = tv_weight * 2 * (
# (tf.nn.l2_loss(inferenceImg[:, 1:, :, :] - inferenceImg[:, :content_shape[1] - 1, :, :]) /
# tv_y_size) +
# (tf.nn.l2_loss(inferenceImg[:, :, 1:, :] - inferenceImg[:, :, :content_shape[2] - 1, :]) /
# tv_x_size))
tv_loss = 0
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer training
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in range(iterations):
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations
and i % print_iterations == 0):
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
print_progress()
if (checkpoint_iterations
and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = inferenceImg.eval()
img_out = vgg.unprocess(best.reshape(content_shape[1:]),
vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(
Image.fromarray(
styled_grayscale_rgb.astype(
np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(
Image.fromarray(original_image.astype(
np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(
Image.fromarray(combined_yuv,
'YCbCr').convert('RGB'))
yield ((None if last_step else i), img_out)
def optimize(content_targets, style_target, content_weight, style_weight,
tv_weight, vgg_path):
mod = len(content_targets) % batch_size
if (mod > 0):
print("Train set has been trimmed slightly..")
content_targets = content_targets[:-mod]
batch_shape = (batch_size, 256, 256, 3)
style_shape = (1, *style_target.shape)
print('batch shape:', batch_shape)
print('style shape:', style_shape)
with tf.Graph().as_default(), tf.Session() as sess:
# Declare placeholders we'll feed into the graph
style_image = tf.placeholder(tf.float32,
shape=style_shape,
name='style_image')
X_content = tf.placeholder(tf.float32,
shape=batch_shape,
name='X_content')
# Precompute content features
start_time = time.time()
content_features = {}
X_content_pre = vgg.preprocess(X_content)
content_net = vgg.net(vgg_path, X_content_pre)
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER]
end_time = time.time()
delta_time = end_time - start_time
print('precompute content features time:', delta_time)
# Precompute style features
start_time = time.time()
style_features = {}
style_pre = np.array([style_target]) # feed
style_image_pre = vgg.preprocess(style_image)
style_net = vgg.net(vgg_path, style_image_pre)
for layer in STYLE_LAYERS:
features = style_net[layer].eval(
feed_dict={style_image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
end_time = time.time()
delta_time = end_time - start_time
print('precompute style features time:', delta_time)
# Build prediction net
preds = transform.net(X_content / 255.0)
preds_pre = vgg.preprocess(preds)
preds_net = vgg.net(vgg_path, preds_pre)
# Compute content loss ?
start_time = time.time()
content_size = _tensor_size(
content_features[CONTENT_LAYER]) * batch_size
assert _tensor_size(content_features[CONTENT_LAYER]) == _tensor_size(
preds_net[CONTENT_LAYER])
content_loss = content_weight * (
2 * tf.nn.l2_loss(preds_net[CONTENT_LAYER] -
content_features[CONTENT_LAYER]) / content_size)
end_time = time.time()
delta_time = end_time - start_time
print('compute content loss time:', delta_time)
# Compute style loss ?
start_time = time.time()
style_losses = []
for style_layer in STYLE_LAYERS:
layer = preds_net[style_layer]
bs, height, width, filters = map(lambda i: i.value,
layer.get_shape())
size = height * width * filters
feats = tf.reshape(layer, (bs, height * width, filters))
feats_T = tf.transpose(feats, perm=[0, 2, 1])
grams = tf.matmul(feats_T, feats) / size
style_gram = style_features[style_layer]
style_losses.append(2 * tf.nn.l2_loss(grams - style_gram) /
style_gram.size)
style_loss = style_weight * functools.reduce(tf.add,
style_losses) / batch_size
end_time = time.time()
delta_time = end_time - start_time
print('compute style loss time:', delta_time)
# Total variation denoising ?
start_time = time.time()
tv_y_size = _tensor_size(preds[:, 1:, :, :])
tv_x_size = _tensor_size(preds[:, :, 1:, :])
y_tv = tf.nn.l2_loss(preds[:, 1:, :, :] -
preds[:, :batch_shape[1] - 1, :, :])
x_tv = tf.nn.l2_loss(preds[:, :, 1:, :] -
preds[:, :, :batch_shape[2] - 1, :])
tv_loss = tv_weight * 2 * (x_tv / tv_x_size +
y_tv / tv_y_size) / batch_size
end_time = time.time()
delta_time = end_time - start_time
print('total variation denoising time:', delta_time)
# Overall loss
start_time = time.time()
all_loss = content_loss + style_loss + tv_loss
end_time = time.time()
delta_time = end_time - start_time
print('compute overall loss time:', delta_time)
# Build train
train = tf.train.AdamOptimizer(learning_rate).minimize(all_loss)
sess.run(tf.global_variables_initializer())
print('Start training...')
start_time = time.time()
num_examples = len(content_targets)
n_batches = num_examples // batch_size
iterations = n_batches * epochs
# For writing training checkpoints.
saver = tf.train.Saver()
for epoch in range(epochs):
for batch in range(n_batches):
iteration = epoch * n_batches + batch + 1
# curr = iteration * batch_size
# step = curr + batch_size
curr = batch * batch_size
step = curr + batch_size
X_batch = np.zeros(batch_shape, dtype=np.float32) # feed
for i, img_p in enumerate(content_targets[curr:step]):
X_batch[i] = get_img(img_p,
(256, 256, 3)).astype(np.float32)
assert X_batch.shape[0] == batch_size
sess.run(train, feed_dict={X_content: X_batch})
to_get = [style_loss, content_loss, tv_loss, all_loss, preds]
if (iteration % display_every_n == 0):
tup = sess.run(to_get, feed_dict={X_content: X_batch})
_style_loss, _content_loss, _tv_loss, _all_loss, _preds = tup
losses = (_style_loss, _content_loss, _tv_loss, _all_loss)
print(
'Iteration {}/{} - style loss: {:.4f}, content loss: {:.4f}, tv loss: {:.4f}, all loss: {:.4f}'
.format(iteration, iterations, *losses))
if (iteration % save_every_n == 0) or (iteration
== iterations):
_all_loss = sess.run(all_loss,
feed_dict={X_content: X_batch})
ckpt = saver.save(
sess,
os.path.join(FLAGS.checkpoint_dir,
"ckpt_i{}".format(iteration)))
print('Epoch {}/{}, Iteration: {}/{}, loss: {}'.format(
epoch, epochs, iteration, iterations, _all_loss))
yield (epoch, iteration, ckpt)
end_time = time.time()
delta_time = end_time - start_time
print('Done! Train total time:', delta_time)
def stylize(content,
style,
initial,
initial_noiseblend,
content_weight=5e0,
content_layer_num=9,
style_weight=5e2,
style_layer_weight=(0.2, 0.2, 0.2, 0.2, 0.2),
tv_weight=1e2,
learning_rate=1e1,
beta1=0.9,
beta2=0.999,
epsilon=1e-8,
preserve_colors=False,
pooling='max',
iterations=1000,
print_iterations=None,
checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1, ) + content.shape
content_features = {}
style_features = {}
style_layers_weights = {}
content_layer = CONTENT_LAYERS[content_layer_num]
for i, style_layer in enumerate(STYLE_LAYERS):
style_layers_weights[style_layer] = style_layer_weight[i]
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
image = tf.placeholder(tf.float32, shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
style_pre = np.array([vgg.preprocess(style, vgg_mean_pixel)])
# compute content features,style features in feedforward mode
with tf.Session() as sess:
content_features[content_layer] = sess.run(
net[content_layer], feed_dict={image: content_pre})
for layer in STYLE_LAYERS:
features = sess.run(net[layer], feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
# make stylized image using backpropogation
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype(np.float32)
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = initial * (1 - initial_noiseblend) + (
tf.random_normal(shape) * 0.256) * initial_noiseblend
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_loss = content_weight * 2 * tf.nn.l2_loss(
net[content_layer] -
content_features[content_layer]) / content_features[content_layer].size
# style loss
style_loss = 0
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[style_layer]
style_loss += style_weight * style_layers_weights[
style_layer] * 2 * tf.nn.l2_loss(gram -
style_gram) / style_gram.size
# total variation denoising
tv_y_size = _tensor_size(image[:, 1:, :, :])
tv_x_size = _tensor_size(image[:, :, 1:, :])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:, 1:, :, :] - image[:, :shape[1] - 1, :, :]) /
tv_y_size) +
(tf.nn.l2_loss(image[:, :, 1:, :] - image[:, :, :shape[2] - 1, :]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
def print_progress():
print(' content loss: %g\n' % content_loss.eval())
print(' style loss: %g\n' % style_loss.eval())
print(' tv loss: %g\n' % tv_loss.eval())
print(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
images = []
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in range(iterations):
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations and i % print_iterations == 0):
print('Iteration %4d/%4d\n' % (i + 1, iterations))
print_progress()
if (checkpoint_iterations
and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
styled_image = np.clip(
vgg.unprocess(image.eval().reshape(shape[1:]),
vgg_mean_pixel), 0, 255)
if this_loss < best_loss:
best_loss = this_loss
best = styled_image
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(
Image.fromarray(styled_grayscale_rgb.astype(
np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(
Image.fromarray(original_image.astype(
np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
styled_image = np.array(
Image.fromarray(combined_yuv, 'YCbCr').convert('RGB'))
plt.figure(figsize=(8, 8))
plt.imshow(styled_image.astype(np.uint8))
plt.axis('off')
plt.show()
images.append(styled_image.astype(np.uint8))
return images, best
style_batch = np.zeros([batch_size, 128, 128])
for i, single in enumerate(style_batch):
style_batch[i] = ss
ss = np.reshape(ss, [-1, 128, 128, 1])
style_features = {}
# precompute style features
with tf.Graph().as_default(), tf.device('/cpu:0'), tf.Session() as sess:
style_image = tf.placeholder(tf.float32,
shape=[None, 128, 128, 1],
name='style_image')
style_image_pre = vgg.preprocess(style_image)
net = vgg.net(vgg_path, style_image_pre)
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={style_image: ss})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
'''
stage 1: generate depth images from joints distribution
'''
gen = Generator
img_gen = gen.generator(X_in_label)
img_gen_trans = tf.reshape(img_gen, [-1, 128, 128])
# loss_generator = tf.reduce_mean(tf.abs(img_gen_trans - X_in_image))
loss_generator = tf.reduce_mean(
def main():
content_path, style_path, width, style_scale = sys.argv[1:]
width = int(width)
style_scale = float(style_scale)
content_image = imread(content_path)
style_image = imread(style_path)
if width > 0:
new_shape = (int(math.floor(float(content_image.shape[0]) /
content_image.shape[1] * width)), width)
content_image = sm.imresize(content_image, new_shape)
if style_scale > 0:
style_image = sm.imresize(style_image, style_scale)
shape = (1,) + content_image.shape
style_shape = (1,) + style_image.shape
content_features = {}
style_features = {}
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net, mean_pixel = vgg.net(VGG_PATH, image)
content_pre = np.array([vgg.preprocess(content_image, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(
feed_dict={image: content_pre})
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shape)
net, _ = vgg.net(VGG_PATH, image)
style_pre = np.array([vgg.preprocess(style_image, mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / (features.size)
style_features[layer] = gram
with tf.Graph().as_default():
noise = np.random.normal(size=shape, scale=np.std(content_image) * 0.1)
init = tf.random_normal(shape) * 256 / 1000
image = tf.Variable(init)
net, _ = vgg.net(VGG_PATH, image)
content_loss = tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER])
style_losses = []
for i in STYLE_LAYERS:
layer = net[i]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / (size)
style_gram = style_features[i]
style_losses.append(tf.nn.l2_loss(gram - style_gram))
style_loss = reduce(tf.add, style_losses) / len(style_losses)
tv_loss = (tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) +
tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]))
loss = ALPHA * content_loss + BETA * style_loss + TV_WEIGHT * tv_loss
train_step = tf.train.AdamOptimizer(LEARNING_RATE).minimize(loss)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(100000):
print 'i = %d' % i
if i % 10 == 0:
print '\tcontent_loss = %15.0f' % content_loss.eval()
print '\tstyle_loss = %15.0f' % style_loss.eval()
print '\ttv_loss = %15.0f' % tv_loss.eval()
print '\tloss = %15.0f' % loss.eval()
imsave('%05d.jpg' % i, vgg.unprocess(
image.eval().reshape(shape[1:]), mean_pixel))
train_step.run()
def main():
'''Search for similar images
Search the style directory for images that closely resemble each image in
the content directory. Save those images in an output directory folder
corresponding to each content image, renamed as their matching rank number.
'''
parser = build_parser()
options = parser.parse_args()
content_files = os.listdir(options.content_dir)
content_images = [
read_img(os.path.join(options.content_dir, f)) for f in content_files
]
# n_content by n_style matrix and list to store the best style images
n_content = len(content_files)
n_total = n_content * options.n_style
best_style_score = np.float('inf') * np.ones((n_content, options.n_style))
best_style_file = np.array([['' for i in range(options.n_style)]
for h in range(n_content)],
dtype=object)
vgg_weights, vgg_mean_pixel = vgg.load_net(options.network)
content_features = [{} for _ in content_images]
for i, c in enumerate(content_images):
with tf.Graph().as_default(), tf.Session() as sess:
image = tf.placeholder('float', shape=(1, ) + c.shape)
net = vgg.net_preloaded(vgg_weights, image, 'max')
content_pre = np.array([vgg.preprocess(c, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[i][layer] = net[layer].eval(
feed_dict={image: content_pre})
final_style_score, final_style_file = search_dir(
content_features, vgg_weights, vgg_mean_pixel, best_style_score,
best_style_file, options.style_dir, options.recurse, options.n_search)
if np.any(np.isinf(final_style_score)):
inf_total = np.sum(np.isinf(final_style_score))
print('%d out of %d style images not found.' % (inf_total, n_total),
'Try rerunning with a smaller n-style.')
raise
sorted_files = final_style_file[np.indices(
(n_content, options.n_style))[0],
final_style_score.argsort()]
format_str = '{0:0>%d}.{1}' % np.ceil(np.log10(n_total))
os.mkdir(options.output_dir)
for i, f in enumerate(content_files):
fname = ''.join(f.split('.')[:-1])
print('Copying style files for %s' % fname)
os.mkdir(os.path.join(options.output_dir, fname))
for j in range(options.n_style):
print(sorted_files[i, j])
img_ext = sorted_files[i, j].split('.')[-1]
shutil.copy(
sorted_files[i, j],
os.path.join(options.output_dir, fname,
format_str.format(j, img_ext)))
def optimize(content_targets,
style_target,
content_weight,
style_weight,
tv_weight,
vgg_path,
epochs=2,
print_iterations=1000,
batch_size=4,
save_path='saver/fns.ckpt',
slow=False,
learning_rate=1e-3,
device='/cpu:0',
debug=False,
total_iterations=-1,
base_model_path=None):
if slow:
batch_size = 1
mod = len(content_targets) % batch_size
if mod > 0:
print("Train set has been trimmed slightly..")
content_targets = content_targets[:-mod]
style_features = {}
batch_shape = (batch_size, 256, 256, 3)
style_shape = (1, ) + style_target.shape
print(style_shape)
# precompute style features
print("Precomputing style features")
sys.stdout.flush()
with tf.Graph().as_default(), tf.device(device), tf.Session(
config=tf.ConfigProto(allow_soft_placement=True)) as sess:
style_image = tf.placeholder(tf.float32,
shape=style_shape,
name='style_image')
style_image_pre = vgg.preprocess(style_image)
net = vgg.net(vgg_path, style_image_pre)
style_pre = np.array([style_target])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={style_image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
with tf.Graph().as_default(), tf.Session() as sess:
X_content = tf.placeholder(tf.float32,
shape=batch_shape,
name="X_content")
X_pre = vgg.preprocess(X_content)
print("Precomputing content features")
sys.stdout.flush()
# precompute content features
content_features = {}
content_net = vgg.net(vgg_path, X_pre)
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER]
if slow:
preds = tf.Variable(
tf.random_normal(X_content.get_shape()) * 0.256)
preds_pre = preds
else:
preds = transform.net(X_content / 255.0)
preds_pre = vgg.preprocess(preds)
print("Building VGG net")
sys.stdout.flush()
net = vgg.net(vgg_path, preds_pre)
content_size = _tensor_size(
content_features[CONTENT_LAYER]) * batch_size
assert _tensor_size(content_features[CONTENT_LAYER]) == _tensor_size(
net[CONTENT_LAYER])
content_loss = content_weight * (
2 * tf.nn.l2_loss(net[CONTENT_LAYER] -
content_features[CONTENT_LAYER]) / content_size)
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
bs, height, width, filters = map(lambda i: i.value,
layer.get_shape())
size = height * width * filters
feats = tf.reshape(layer, (bs, height * width, filters))
feats_T = tf.transpose(feats, perm=[0, 2, 1])
# see https://github.com/tensorflow/tensorflow/issues/6560
grams = tf.matmul(feats_T, feats) / size
style_gram = style_features[style_layer]
style_losses.append(2 * tf.nn.l2_loss(grams - style_gram) /
style_gram.size)
style_loss = style_weight * reduce(tf.add, style_losses) / batch_size
# total variation denoising
tv_y_size = _tensor_size(preds[:, 1:, :, :])
tv_x_size = _tensor_size(preds[:, :, 1:, :])
y_tv = tf.nn.l2_loss(preds[:, 1:, :, :] -
preds[:, :batch_shape[1] - 1, :, :])
x_tv = tf.nn.l2_loss(preds[:, :, 1:, :] -
preds[:, :, :batch_shape[2] - 1, :])
tv_loss = tv_weight * 2 * (x_tv / tv_x_size +
y_tv / tv_y_size) / batch_size
loss = content_loss + style_loss + tv_loss
# overall loss
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
sess.run(tf.initialize_all_variables())
# If base model file is present, load that in to the session
if base_model_path:
saver = tf.train.Saver()
if os.path.isdir(base_model_path):
ckpt = tf.train.get_checkpoint_state(base_model_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
raise Exception("No checkpoint found...")
else:
saver.restore(sess, base_model_path)
import random
uid = random.randint(1, 100)
print("UID: %s" % uid)
sys.stdout.flush()
for epoch in range(epochs):
num_examples = len(content_targets)
print("number of examples: %s" % num_examples)
sys.stdout.flush()
iterations = 0
while iterations * batch_size < num_examples:
print("Current iteration : %s" % iterations)
sys.stdout.flush()
start_time = time.time()
curr = iterations * batch_size
step = curr + batch_size
X_batch = np.zeros(batch_shape, dtype=np.float32)
for j, img_p in enumerate(content_targets[curr:step]):
X_batch[j] = get_img(img_p,
(256, 256, 3)).astype(np.float32)
iterations += 1
assert X_batch.shape[0] == batch_size
feed_dict = {X_content: X_batch}
train_step.run(feed_dict=feed_dict)
end_time = time.time()
delta_time = end_time - start_time
if debug:
print("UID: %s, batch time: %s" % (uid, delta_time))
is_print_iter = int(iterations) % print_iterations == 0
if slow:
is_print_iter = epoch % print_iterations == 0
is_last = False
if epoch == epochs - 1 and iterations * batch_size >= num_examples:
is_last = True
if total_iterations > 0 and iterations >= total_iterations:
is_last = True
should_print = is_print_iter or is_last
if should_print:
to_get = [style_loss, content_loss, tv_loss, loss, preds]
test_feed_dict = {X_content: X_batch}
tup = sess.run(to_get, feed_dict=test_feed_dict)
_style_loss, _content_loss, _tv_loss, _loss, _preds = tup
losses = (_style_loss, _content_loss, _tv_loss, _loss)
if slow:
_preds = vgg.unprocess(_preds)
else:
saver = tf.train.Saver()
res = saver.save(sess, save_path)
yield (_preds, losses, iterations, epoch)
if is_last:
break
def optimize(content_targets, style_target, content_weight, style_weight,
tv_weight, vgg_path, epochs=2, print_iterations=1000,
batch_size=4, save_path='saver/fns.ckpt', slow=False,
learning_rate=1e-3, debug=False, max_sample=4000):
if slow:
batch_size = 1
mod = len(content_targets) % batch_size
if mod > 0:
print("Train set has been trimmed slightly..")
content_targets = content_targets[:-mod]
if len(content_targets) > max_sample:
content_targets = content_targets[:max_sample]
style_features = {}
batch_shape = (batch_size,256,256,3)
style_shape = (1,) + style_target.shape
print(style_shape)
# precompute style features
with tf.Graph().as_default(), tf.device('/cpu:0'), tf.Session() as sess:
style_image = tf.placeholder(tf.float32, shape=style_shape, name='style_image')
style_image_pre = vgg.preprocess(style_image)
net = vgg.net(vgg_path, style_image_pre)
style_pre = np.array([style_target])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={style_image:style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
with tf.Graph().as_default(), tf.Session() as sess:
X_content = tf.placeholder(tf.float32, shape=batch_shape, name="X_content")
X_pre = vgg.preprocess(X_content)
# precompute content features
content_features = {}
content_net = vgg.net(vgg_path, X_pre)
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER]
if slow:
preds = tf.Variable(
tf.random_normal(X_content.get_shape()) * 0.256
)
preds_pre = preds
else:
preds = transform.net(X_content/255.0)
preds_pre = vgg.preprocess(preds)
net = vgg.net(vgg_path, preds_pre)
content_size = _tensor_size(content_features[CONTENT_LAYER])*batch_size
assert _tensor_size(content_features[CONTENT_LAYER]) == _tensor_size(net[CONTENT_LAYER])
content_loss = content_weight * (2 * tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) / content_size
)
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
bs, height, width, filters = map(lambda i:i.value,layer.get_shape())
size = height * width * filters
feats = tf.reshape(layer, (bs, height * width, filters))
feats_T = tf.transpose(feats, perm=[0,2,1])
grams = tf.matmul(feats_T, feats) / size
style_gram = style_features[style_layer]
style_losses.append(2 * tf.nn.l2_loss(grams - style_gram)/style_gram.size)
style_loss = style_weight * functools.reduce(tf.add, style_losses) / batch_size
# total variation denoising
tv_y_size = _tensor_size(preds[:,1:,:,:])
tv_x_size = _tensor_size(preds[:,:,1:,:])
y_tv = tf.nn.l2_loss(preds[:,1:,:,:] - preds[:,:batch_shape[1]-1,:,:])
x_tv = tf.nn.l2_loss(preds[:,:,1:,:] - preds[:,:,:batch_shape[2]-1,:])
tv_loss = tv_weight*2*(x_tv/tv_x_size + y_tv/tv_y_size)/batch_size
loss = content_loss + style_loss + tv_loss
# overall loss
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
sess.run(tf.initialize_all_variables())
import random
uid = random.randint(1, 100)
print("UID: %s" % uid)
for epoch in range(epochs):
num_examples = len(content_targets)
iterations = 0
while iterations * batch_size < num_examples:
start_time = time.time()
curr = iterations * batch_size
step = curr + batch_size
X_batch = np.zeros(batch_shape, dtype=np.float32)
for j, img_p in enumerate(content_targets[curr:step]):
X_batch[j] = get_img(img_p, (256,256,3)).astype(np.float32)
iterations += 1
assert X_batch.shape[0] == batch_size
feed_dict = {
X_content:X_batch
}
train_step.run(feed_dict=feed_dict)
end_time = time.time()
delta_time = end_time - start_time
if debug:
print("UID: %s, batch time: %s" % (uid, delta_time))
is_print_iter = int(iterations) % print_iterations == 0
if slow:
is_print_iter = epoch % print_iterations == 0
is_last = epoch == epochs - 1 and iterations * batch_size >= num_examples
should_print = is_print_iter or is_last
if should_print:
to_get = [style_loss, content_loss, tv_loss, loss, preds]
test_feed_dict = {
X_content:X_batch
}
tup = sess.run(to_get, feed_dict = test_feed_dict)
_style_loss,_content_loss,_tv_loss,_loss,_preds = tup
losses = (_style_loss, _content_loss, _tv_loss, _loss)
if slow:
_preds = vgg.unprocess(_preds)
else:
saver = tf.train.Saver()
res = saver.save(sess, save_path)
yield(_preds, losses, iterations, epoch)
def main():
# This will print all array values in full
np.set_printoptions(threshold=np.nan)
parser = build_parser()
options = parser.parse_args()
if not os.path.isfile(options.network):
parser.error(
"Network %s does not exist. (Did you forget to download it?)" %
options.network)
# Load the vgg weights in advance
vgg_weights, vgg_mean_pixel = vgg.load_net(options.network)
content_image = imread(options.content)
# Jacob: moved this here since the same image features will be used for each style image
content_features = {}
g = tf.Graph()
shape = (1, ) + content_image.shape
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, options.pooling)
content_pre = np.array([vgg.preprocess(content_image, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(
feed_dict={image: content_pre})
print("READY")
sys.stdout.flush(
) # Make sure Java can sense this output before Python blocks waiting for input
count = 0
#for style in style_images: # loop through separate style inputs individually
for line in sys.stdin:
# Assumes a single line of input will be a json for one image
style = jsonimread(line)
width = options.width
if width is not None:
new_shape = (int(
math.floor(
float(content_image.shape[0]) / content_image.shape[1] *
width)), width)
content_image = scipy.misc.imresize(content_image, new_shape)
target_shape = content_image.shape
# This batch of code was in a loop for each style input before
style_scale = STYLE_SCALE
if options.style_scales is not None:
style_scale = options.style_scales[i]
style = scipy.misc.imresize(
style, style_scale * target_shape[1] / style.shape[1])
# Removed code for blanding between multiple styles
style_blend_weights = [1.0]
initial = options.initial
if initial is not None:
initial = scipy.misc.imresize(imread(initial),
content_image.shape[:2])
# Initial guess is specified, but not noiseblend - no noise should be blended
if options.initial_noiseblend is None:
options.initial_noiseblend = 0.0
else:
# Neither inital, nor noiseblend is provided, falling back to random generated initial guess
if options.initial_noiseblend is None:
options.initial_noiseblend = 1.0
if options.initial_noiseblend < 1.0:
initial = content_image
if options.checkpoint_output and "%s" not in options.checkpoint_output:
parser.error("To save intermediate images, the checkpoint output "
"parameter must contain `%s` (e.g. `foo%s.jpg`)")
for iteration, image in stylize(
network=options.network,
initial=initial,
initial_noiseblend=options.initial_noiseblend,
content=content_image,
styles=[style
], # Changed this to be a list of only one style image
preserve_colors=options.preserve_colors,
iterations=options.iterations,
content_weight=options.content_weight,
content_weight_blend=options.content_weight_blend,
style_weight=options.style_weight,
style_layer_weight_exp=options.style_layer_weight_exp,
style_blend_weights=style_blend_weights,
tv_weight=options.tv_weight,
learning_rate=options.learning_rate,
beta1=options.beta1,
beta2=options.beta2,
epsilon=options.epsilon,
pooling=options.pooling,
print_iterations=options.print_iterations,
checkpoint_iterations=options.checkpoint_iterations,
# These vgg settings are now loaded only once
vgg_weights=vgg_weights,
vgg_mean_pixel=vgg_mean_pixel,
content_features=content_features):
output_file = None
combined_rgb = image
if iteration is not None:
if options.checkpoint_output:
output_file = options.checkpoint_output % iteration
else:
# Change final output files to simply be numbered
output_file = "%d.JPG" % count
count = count + 1
if output_file:
# No longer save image to file
#imsave(output_file, combined_rgb)
# Output json String
print(json.dumps(combined_rgb.tolist()))
sys.stdout.flush(
) # Make sure Java can sense this output before Python blocks waiting for input
print("DONE")
def stylize(network,
initial,
initial_noiseblend,
content,
styles,
luminance_transfer,
iterations,
content_weight,
content_weight_blend,
style_weight,
style_layer_weight_exp,
style_blend_weights,
tv_weight,
learning_rate,
beta1,
beta2,
epsilon,
pooling,
print_iterations=None,
checkpoint_iterations=None):
"""
This function yields tuples (iteration, image).
`iteration` is None if this is the final image (the last iteration).
Other tuples are yielded every `checkpoint_iterations` iterations.
"""
shape = (1, ) + content.shape
style_shapes = [(1, ) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, pooling)
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (
tf.random_normal(shape) *
0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_layers_weights = {
'relu4_2': content_weight_blend,
'relu5_2': 1.0 - content_weight_blend
}
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(
content_layers_weights[content_layer] * content_weight *
(2 * tf.nn.l2_loss(net[content_layer] -
content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value,
layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(style_layers_weights[style_layer] * 2 *
tf.nn.l2_loss(gram - style_gram) /
style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(
tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:, 1:, :, :])
tv_x_size = _tensor_size(image[:, :, 1:, :])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:, 1:, :, :] - image[:, :shape[1] - 1, :, :])
/ tv_y_size) +
(tf.nn.l2_loss(image[:, :, 1:, :] - image[:, :, :shape[2] - 1, :])
/ tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in tqdm(range(iterations)):
#stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations
and i % print_iterations == 0):
print_progress()
if (checkpoint_iterations
and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]),
vgg_mean_pixel)
if luminance_transfer and luminance_transfer == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized image into YUV (YCbCr)
# 2. Convert original image into YUV (YCbCr)
# 3. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 4. Convert recombined image from YUV back to RGB
# 1
styled_grayscale_yuv = np.array(
Image.fromarray(styled_image.astype(
np.uint8)).convert('YCbCr'))
# 2
original_yuv = np.array(
Image.fromarray(original_image.astype(
np.uint8)).convert('YCbCr'))
# 3
h, w, _ = original_image.shape
combined_yuv = np.empty((h, w, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 4
img_out = np.array(
Image.fromarray(combined_yuv,
'YCbCr').convert('RGB'))
yield ((None if last_step else i), img_out)
def main():
global options, device
# Get the ENV context
script_dir = os.path.dirname(__file__)
env = os.environ.copy()
# Set the input folder
input_dir = os.path.expanduser(options.input_dir) if options.input_dir \
else os.path.join(script_dir, '..', 'data')
vgg_path = os.path.join(input_dir, 'vgg', 'imagenet-vgg-verydeep-19.mat')
coco_dir = os.path.join(input_dir, 'train')
if not os.path.isdir(input_dir):
fail('Failed to find the input folder at ' + input_dir)
if not os.path.isfile(vgg_path):
error('Failed to find the VGG model file at ' + vgg_path)
fail(
'Please download it from http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat'
)
if not os.path.isdir(coco_dir):
error('Failed to find the COCO 2014 training images in ' + coco_dir)
fail(
'Plese download it from http://images.cocodataset.org/zips/train2014.zip'
)
# Set the output folder
output_dir = os.path.expanduser(options.output_dir) if options.output_dir \
else env.get('OUTPUT_DIR', os.path.join(script_dir, '..', 'output'))
model_dir = os.path.join(output_dir, 'checkpoint')
export_dir = os.path.join(output_dir, 'savedmodel')
if os.path.isdir(output_dir):
if not os.path.isdir(model_dir):
info('Creating a folder to store checkpoint at ' + model_dir)
os.makedirs(model_dir)
if os.path.isdir(export_dir):
info('Deleting the folder containing SavedModel at ' + export_dir)
shutil.rmtree(export_dir)
else:
info('Creating a folder to store checkpoint at ' + model_dir)
os.makedirs(model_dir)
# Set the TensorBoard folder
log_dir = os.path.expanduser(options.log_dir) if options.log_dir \
else env.get('LOG_DIR', os.path.join(script_dir, '..', 'log'))
if not os.path.isdir(log_dir):
info('Creating a folder to store TensorBoard events at ' + log_dir)
os.makedirs(log_dir)
# Set the style image path
style_path = os.path.expanduser(options.style_image) if os.path.isfile(options.style_image) \
else os.path.join(input_dir, 'style_images', options.style_image)
style_name = os.path.basename(os.path.splitext(style_path)[0])
ckpt_path = os.path.join(model_dir, style_name + '.ckpt')
if not os.path.isfile(style_path):
fail('Failed to find the style image at ' + style_path)
# Set hyper parameters
batch_size = options.batch_size
epochs = options.epoch
lr = options.lr
lambda_tv = options.lambda_tv
lambda_feat = options.lambda_feat
lambda_style = options.lambda_style
# Print parsed arguments
info('--------- Training parameters -------->')
info('Style image path: ' + style_path)
info('VGG model path: ' + vgg_path)
info('Training image dir: ' + coco_dir)
info('Checkpoint path: ' + ckpt_path)
info('TensorBoard log dir: ' + log_dir)
info('Training device: ' + device)
info('Batch size: %d' % batch_size)
info('Epoch count: %d' % epochs)
info('Learning rate: ' + str(lr))
info('Lambda tv: ' + str(lambda_tv))
info('Lambda feat: ' + str(lambda_feat))
info('Lambda style: ' + str(lambda_style))
info('<-------- Training parameters ---------')
# COCO images to train
content_targets = list_jpgs(coco_dir)
if len(content_targets) % batch_size != 0:
content_targets = content_targets[:-(len(content_targets) %
batch_size)]
info('Total training data size: %d' % len(content_targets))
# Image shape
image_shape = (224, 224, 3)
batch_shape = (batch_size, ) + image_shape
# Style target
style_target = read_img(style_path)
style_shape = (1, ) + style_target.shape
with tf.device(device), tf.Session() as sess:
# Compute gram maxtrix of style target
style_image = tf.placeholder(tf.float32,
shape=style_shape,
name='style_image')
vggstyletarget = vgg.net(vgg_path, vgg.preprocess(style_image))
style_vgg = vgg.get_style_vgg(vggstyletarget, style_image,
np.array([style_target]))
# Content target feature
content_vgg = {}
inputs = tf.placeholder(tf.float32, shape=batch_shape, name='inputs')
content_net = vgg.net(vgg_path, vgg.preprocess(inputs))
content_vgg['relu4_2'] = content_net['relu4_2']
# Feature after transformation
outputs = stylenet.net(inputs / 255.0)
vggoutputs = vgg.net(vgg_path, vgg.preprocess(outputs))
# Compute feature loss
loss_f = options.lambda_feat * vgg.total_content_loss(
vggoutputs, content_vgg, batch_size)
# Compute style loss
loss_s = options.lambda_style * vgg.total_style_loss(
vggoutputs, style_vgg, batch_size)
# Total variation denoising
loss_tv = options.lambda_tv * vgg.total_variation_regularization(
outputs, batch_size, batch_shape)
# Total loss
total_loss = loss_f + loss_s + loss_tv
train_step = tf.train.AdamOptimizer(options.lr).minimize(total_loss)
# Create summary
tf.summary.scalar('loss', total_loss)
merged = tf.summary.merge_all()
# Used to save model
saver = tf.train.Saver()
builder = tf.saved_model.builder.SavedModelBuilder(export_dir)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
# Restore checkpoint if available
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(model_dir)
if ckpt and ckpt.model_checkpoint_path:
info('Restoring from ' + ckpt.model_checkpoint_path)
saver.restore(sess, ckpt.model_checkpoint_path)
# Write the graph
writer = tf.summary.FileWriter(log_dir, sess.graph)
# Start to train
total_step = 0
for epoch in range(epochs):
info('epoch: %d' % epoch)
step = 0
while step * batch_size < len(content_targets):
time_start = time.time()
# Load one batch
batch = np.zeros(batch_shape, dtype=np.float32)
for i, img in enumerate(
content_targets[step * batch_size:(step + 1) *
batch_size]):
batch[i] = read_img(img, image_shape).astype(
np.float32) # (224,224,3)
# Proceed one step
step += 1
total_step += 1
_, loss, summary = sess.run([train_step, total_loss, merged],
feed_dict={inputs: batch})
time_elapse = time.time() - time_start
if total_step % 5 == 0:
info('[step {}] elapse time: {} loss: {}'.format(
total_step, time_elapse, loss))
writer.add_summary(summary, total_step)
# Write checkpoint
if total_step % 2000 == 0:
info('Saving checkpoint to ' + ckpt_path)
saver.save(sess, ckpt_path, global_step=total_step)
info('Exporting SavedModel to ' + export_dir)
serving_signatures = {
'Transfer': #tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
tf.saved_model.signature_def_utils.predict_signature_def(
{ tf.saved_model.signature_constants.PREDICT_INPUTS: inputs },
{ tf.saved_model.signature_constants.PREDICT_OUTPUTS: outputs }
)
}
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map=serving_signatures,
clear_devices=True)
builder.save()
def stylize(network, initial, initial_noiseblend, content, styles, preserve_colors, iterations,
content_weight, content_weight_blend, style_weight, style_layer_weight_exp, style_blend_weights, tv_weight,
learning_rate, beta1, beta2, epsilon, pooling,
print_iterations=None, checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
vgg_weights, vgg_mean_pixel = vgg.load_net(network)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_preloaded(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net = vgg.net_preloaded(vgg_weights, image, pooling)
style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = (initial) * initial_content_noise_coeff + (tf.random_normal(shape) * 0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_preloaded(vgg_weights, image, pooling)
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(content_layers_weights[content_layer] * content_weight * (2 * tf.nn.l2_loss(
net[content_layer] - content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(style_layers_weights[style_layer] * 2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2, epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
iteration_times = []
start = time.time()
for i in range(iterations):
iteration_start = time.time()
if i > 0:
elapsed = time.time() - start
# take average of last couple steps to get time per iteration
remaining = np.mean(iteration_times[-10:]) * (iterations - i)
stderr.write('Iteration %4d/%4d (%s elapsed, %s remaining)\n' % (
i + 1,
iterations,
hms(elapsed),
hms(remaining)
))
else:
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations and i % print_iterations == 0):
print_progress()
if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(Image.fromarray(original_image.astype(np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(Image.fromarray(combined_yuv, 'YCbCr').convert('RGB'))
yield (
(None if last_step else i),
img_out
)
iteration_end = time.time()
iteration_times.append(iteration_end - iteration_start)
def optimize(content_targets, style_target, content_weight, style_weight,
tv_weight, vgg_path, use_IN, epochs=2, print_iterations=1000,
batch_size=4, save_path='checkpoints/fast_style_transfer.ckpt', slow=False,
learning_rate=1e-3, debug=False):
if slow:
batch_size = 1
# content_target is a list of files, 4-D size, so this is about the batch size here.
# If using only one content image, then mod here is 0.
mod = len(content_targets) % batch_size
if mod > 0:
print("Train set has been trimmed slightly...")
content_targets = content_targets[:-mod]
# training image get to be 256 x 256 because of get_img resize,
# it then get into tensorflow graph from Adam optimizer feed_dict.
batch_shape = (batch_size, 256, 256, 3)
style_shape = (1,) + style_target.shape # add 1 in the front for batch size, 4-D.
print(f"batch_shape of the content image is: {batch_shape}")
print(f"style_shape of the style image is: {style_shape}")
### Graph Construction ###
# vgg won't be trained, because in vgg.py the weights are loaded through that matlab file.
# computed vgg style features in gram matrices
# tf.device('/cpu:0')
config = v1.ConfigProto()
config.gpu_options.allow_growth = True
style_features = {}
with tf.Graph().as_default(), v1.Session(config=config) as sess:
style_image = v1.placeholder(tf.float32, shape=style_shape, name='style_image') # 4-D placeholder for feed_dict
vgg_style_net = vgg.net(vgg_path, vgg.preprocess(style_image)) # extract feature volume
np_style_target = np.array([style_target]) # a 3-D numpy array for feed_dict's input
for layer in STYLE_LAYERS:
# vgg_style_net[layer] is a tf.Tensor returned by tf.nn.relu,
# eval at that layer, by running forward to that vgg layer or entire network.
features = vgg_style_net[layer].eval(feed_dict={style_image:np_style_target}) # extract a fVol value
features = np.reshape(features, (-1, features.shape[3])) # (N*H*W, C)
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
# computed vgg content feature map and both losses
with tf.Graph().as_default(), v1.Session(config=config) as sess:
X_content = v1.placeholder(tf.float32, shape=batch_shape, name="X_content") # 4-D
vgg_content_net = vgg.net(vgg_path, vgg.preprocess(X_content)) # run ground truth image through the pre-trained model
# noisy prediction image runs through feed forward conv net, then
# run through vgg to extract feature volume predicitons
if slow:
preds = tf.Variable(
tf.random.normal(X_content.get_shape()) * 0.256
)
preds_pre = preds
else:
preds = transform.net(X_content/255.0, use_IN) # run through the style feed forward network. why need to normalize pixel to 0-1?
net = vgg.net(vgg_path, vgg.preprocess(preds)) # run generated image through the pre-trained model
# _tensor_size is a reduce function only count from [1:],
# so it doesn't have batch_size information.
content_size = _tensor_size(vgg_content_net[CONTENT_LAYER]) * batch_size
vgg_content_net_size = _tensor_size(vgg_content_net[CONTENT_LAYER])
vgg_transform_content_net_size = _tensor_size(net[CONTENT_LAYER])
# print(f"vgg_content_net_size is {vgg_content_net_size}")
# print(vgg_content_net[CONTENT_LAYER])
# print(f"vgg_transform_content_net_size is {vgg_transform_content_net_size}")
# print(net[CONTENT_LAYER])
assert vgg_content_net_size == vgg_transform_content_net_size
# define loss functions
# content loss
content_l2_loss = 2 * tf.nn.l2_loss(net[CONTENT_LAYER] - vgg_content_net[CONTENT_LAYER])
content_loss = content_weight * (content_l2_loss / content_size)
# style loss
style_l2_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
N, H, W, C = map(lambda i : i, layer.get_shape())
feats = tf.reshape(layer, (N, H*W, C)) # N, HW, C
feats_T = tf.transpose(feats, perm=[0, 2, 1]) # N, C, HW
pred_gram = tf.matmul(feats_T, feats) / (H * W * C)
true_gram = style_features[style_layer] # numpy array
style_l2_loss = 2 * tf.nn.l2_loss(pred_gram - true_gram)
style_l2_losses.append(style_l2_loss / true_gram.size)
style_loss = style_weight * functools.reduce(tf.add, style_l2_losses) / batch_size
# total variation denoising regularization loss
# test if not needed in NN conv case and mirror padding
# tv_y_size = _tensor_size(preds[:,1:,:,:])
# tv_x_size = _tensor_size(preds[:,:,1:,:])
# # N, H, W, C
# y_tv = 2 * tf.nn.l2_loss(preds[:, 1:, :, :] - preds[:, :batch_shape[1]-1, :, :]) # H, down - up
# x_tv = 2 * tf.nn.l2_loss(preds[:, :, 1:, :] - preds[:, :, :batch_shape[2]-1, :]) # W, right - left
# tv_loss = tv_weight * (x_tv/tv_x_size + y_tv/tv_y_size) / batch_size
# total loss
# total_loss = content_loss + style_loss + tv_loss
total_loss = content_loss + style_loss
# train the feed forward net, and save weights to a checkpoint.
import random
uid = random.randint(1, 100)
print("This random UID is: %s" % uid)
optimizer = v1.train.AdamOptimizer(learning_rate).minimize(total_loss)
sess.run(v1.global_variables_initializer())
for epoch in range(epochs): # epoch loop
iterations = 0
num_examples = len(content_targets) # COCO train2014 ~20000 images
while iterations * batch_size < num_examples: # batch loop
# start training a batch
start_time = time.time()
X_batch = np.zeros(batch_shape, dtype=np.float32)
start = iterations * batch_size
end = iterations * batch_size + batch_size
for i, img_p in enumerate(content_targets[start:end]): # img_p is a coco images
X_batch[i] = get_img(img_p, (256,256,3)).astype(np.float32) # resize to 256 x 256
optimizer.run(feed_dict={X_content:X_batch})
end_time = time.time()
# end training a batch
# update training information
iterations += 1
is_print_iter = int(iterations) % print_iterations == 0
is_last_train = epoch == epochs - 1 and iterations * batch_size >= num_examples
if slow:
is_print_iter = epoch % print_iterations == 0
if debug:
print("UID: %s, batch training time: %s" % (uid, end_time - start_time))
# monitor the training losses
if is_print_iter or is_last_train:
_style_loss, _content_loss, _total_loss, _preds = \
sess.run([style_loss, content_loss, total_loss, preds],
feed_dict={X_content:X_batch})
losses = (_style_loss, _content_loss, _total_loss)
generated_image = _preds
if slow:
generated_image = vgg.unprocess(generated_image)
else:
res = v1.train.Saver().save(sess, save_path)
print("yield")
yield(generated_image, losses, iterations, epoch)
def stylize(network, initial, content, styles, iterations,
content_weight, style_weight, style_blend_weights, tv_weight,
learning_rate, print_iterations=None, checkpoint_iterations=None):
"""
Stylize images.
This function yields tuples (iteration, image); `iteration` is None
if this is the final image (the last iteration). Other tuples are yielded
every `checkpoint_iterations` iterations.
:rtype: iterator[tuple[int|None,image]]
"""
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net, mean_pixel = vgg.net(network, image)
content_pre = np.array([vgg.preprocess(content, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net, _ = vgg.net(network, image)
style_pre = np.array([vgg.preprocess(styles[i], mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, mean_pixel)])
initial = initial.astype('float32')
image = tf.Variable(initial)
net, _ = vgg.net(network, image)
# content loss
content_loss = content_weight * (2 * tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) /
content_features[CONTENT_LAYER].size)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
def print_progress(i, last=False):
global timenow
stderr.write('Iteration %d/%d, time: %dms\n' % (i + 1, iterations, current_milli_time() - timenow))
timenow = current_milli_time()
if last or (print_iterations and i % print_iterations == 0):
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(iterations):
last_step = (i == iterations - 1)
print_progress(i, last=last_step)
train_step.run()
if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
yield (
(None if last_step else i),
vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
)
def model_neural_style(pre_train_vgg_path,
content_image,
style_images,
content_weight=5e0,
content_weight_blend=1.0,
style_weight=5e2,
style_layer_weight_exp=1.0,
pooling='',
initial=None,
initial_noiseblend=1.0,
tv_weight=1e2,
learning_rate=1e1,
beta1=0.9,
beta2=0.999,
epsilon=1e-08,
print_iterations=None,
iterations=500,
checkpoint_iterations=50,
preserve_colors=None):
print "++++++++++++++++++++"
# input shape of model
shape = (1, ) + content_image.shape
style_images_shapes = [(1, ) + style_image.shape
for style_image in style_images]
content_features = {}
style_features = [{} for _ in style_images]
# load the weights of pretrained vgg model
vgg_weights, vgg_mean_pixel = vgg.load_weights(pre_train_vgg_path)
layer_weight = 1.0
style_layers_weights = {}
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] = layer_weight
layer_weight *= style_layer_weight_exp
# normalize style layer weights
layer_weights_sum = 0
for style_layer in STYLE_LAYERS:
layer_weights_sum += style_layers_weights[style_layer]
for style_layer in STYLE_LAYERS:
style_layers_weights[style_layer] /= layer_weights_sum
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net = vgg.net_infer(vgg_weights, image, pooling)
content_pre = np.array([vgg.preprocess(content_image, vgg_mean_pixel)])
for layer in CONTENT_LAYERS:
content_features[layer] = net[layer].eval(
feed_dict={image: content_pre})
# # for debug
# for layer in CONTENT_LAYERS:
# item = content_features[layer]
# item = item.reshape(item.shape[1], item.shape[2], item.shape[3])
# item_for_plot = []
# for i in range(item.shape[2]):
# item_for_plot.append(item[:, :, i])
#
# tools.show_images(item_for_plot[::8], cols=8)
# compute style features in feedforward mode
# compute styles features in feedforward mode
for i in range(len(style_images)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_images_shapes[i])
net = vgg.net_infer(vgg_weights, image, pooling)
style_pre = np.array(
[vgg.preprocess(style_images[i], vgg_mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[i][layer] = gram
initial_content_noise_coeff = 1.0 - initial_noiseblend
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape,
scale=np.std(content_image) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
initial = initial.astype('float32')
noise = np.random.normal(size=shape,
scale=np.std(content_image) * 0.1)
initial = (initial) * initial_content_noise_coeff + (
tf.random_normal(shape) *
0.256) * (1.0 - initial_content_noise_coeff)
image = tf.Variable(initial)
net = vgg.net_infer(vgg_weights, image, pooling)
# content loss
content_layers_weights = {}
content_layers_weights['relu4_2'] = content_weight_blend
content_layers_weights['relu5_2'] = 1.0 - content_weight_blend
content_loss = 0
content_losses = []
for content_layer in CONTENT_LAYERS:
content_losses.append(
content_layers_weights[content_layer] * content_weight *
(2 * tf.nn.l2_loss(net[content_layer] -
content_features[content_layer]) /
content_features[content_layer].size))
content_loss += reduce(tf.add, content_losses)
# style loss
style_loss = 0
for i in range(len(style_images)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value,
layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(style_layers_weights[style_layer] * 2 *
tf.nn.l2_loss(gram - style_gram) /
style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(
tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:, 1:, :, :])
tv_x_size = _tensor_size(image[:, :, 1:, :])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:, 1:, :, :] - image[:, :shape[1] - 1, :, :])
/ tv_y_size) +
(tf.nn.l2_loss(image[:, :, 1:, :] - image[:, :, :shape[2] - 1, :])
/ tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2,
epsilon).minimize(loss)
def print_progress():
stderr.write(' content loss: %g\n' % content_loss.eval())
stderr.write(' style loss: %g\n' % style_loss.eval())
stderr.write(' tv loss: %g\n' % tv_loss.eval())
stderr.write(' total loss: %g\n' % loss.eval())
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
stderr.write('Optimization started...\n')
if (print_iterations and print_iterations != 0):
print_progress()
for i in range(iterations):
stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations))
train_step.run()
last_step = (i == iterations - 1)
if last_step or (print_iterations
and i % print_iterations == 0):
print_progress()
if (checkpoint_iterations
and i % checkpoint_iterations == 0) or last_step:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
img_out = vgg.unprocess(best.reshape(shape[1:]),
vgg_mean_pixel)
if preserve_colors and preserve_colors == True:
original_image = np.clip(content_image, 0, 255)
styled_image = np.clip(img_out, 0, 255)
# Luminosity transfer steps:
# 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
# 2. Convert stylized grayscale into YUV (YCbCr)
# 3. Convert original image into YUV (YCbCr)
# 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
# 5. Convert recombined image from YUV back to RGB
# 1
styled_grayscale = rgb2gray(styled_image)
styled_grayscale_rgb = gray2rgb(styled_grayscale)
# 2
styled_grayscale_yuv = np.array(
Image.fromarray(
styled_grayscale_rgb.astype(
np.uint8)).convert('YCbCr'))
# 3
original_yuv = np.array(
Image.fromarray(original_image.astype(
np.uint8)).convert('YCbCr'))
# 4
w, h, _ = original_image.shape
combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
combined_yuv[..., 1] = original_yuv[..., 1]
combined_yuv[..., 2] = original_yuv[..., 2]
# 5
img_out = np.array(
Image.fromarray(combined_yuv,
'YCbCr').convert('RGB'))
yield ((None if last_step else i), img_out)
def stylize(network, initial, content, styles, iterations,
content_weight, style_weight, style_blend_weights, tv_weight,
learning_rate, print_iterations=None, checkpoint_iterations=None):
shape = (1,) + content.shape
style_shapes = [(1,) + style.shape for style in styles]
content_features = {}
style_features = [{} for _ in styles]
# compute content features in feedforward mode
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
net, mean_pixel = vgg.net(network, image)
content_pre = np.array([vgg.preprocess(content, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(
feed_dict={image: content_pre})
# compute style features in feedforward mode
for i in range(len(styles)):
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
image = tf.placeholder('float', shape=style_shapes[i])
net, _ = vgg.net(network, image)
style_pre = np.array([vgg.preprocess(styles[i], mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
print 'Initial feature shape: ', features.shape
features = np.reshape(features, (-1, features.shape[3]))
#mask = np.zeros_like(features)
#mask[:49664/2, :] = 1
#print 'Mask shape', mask.shape
print 'Final features shape', features.shape
#features = features*mask
gram = np.matmul(features.T, features) / features.size
print 'Gram matrix shape: ', gram.shape
style_features[i][layer] = gram
#sys.exit()
# make stylized image using backpropogation
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 0.256
else:
initial = np.array([vgg.preprocess(initial, mean_pixel)])
initial = initial.astype('float32')
image = tf.Variable(initial)
net, _ = vgg.net(network, image)
# content loss
content_loss = content_weight * (2 * tf.nn.l2_loss(
net[CONTENT_LAYER] - content_features[CONTENT_LAYER]) /
content_features[CONTENT_LAYER].size)
# style loss
style_loss = 0
for i in range(len(styles)):
style_losses = []
for style_layer in STYLE_LAYERS:
layer = net[style_layer]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
print 'Height, width, number', height, width, number
size = height * width * number
feats = tf.reshape(layer, (-1, number))
#print tf.shape(feats).as_list()
print 'Height', height
print 'Weight', width
print 'Number', number
print 'Style features shape', style_features[i][style_layer].shape
print style_layer
if style_layer == 'relu2_1':
mask = np.zeros((height*width, number), dtype=np.float32)
temp = imread('emma/emma_test_mask.jpg').astype(np.float32)
c = temp.reshape(height,2,width,2)
temp = c.max(axis=1).max(axis=2)
print temp.shape
maskt = np.reshape(temp, (height*width,))
maskt = maskt > 100
for d in xrange(number):
mask[:,d] = maskt
print 'Mask shape', mask.shape
#b = mask.reshape(height*width*2, 2, number/2,2)
#mask = b.max(axis=1).max(axis=2)
#print 'New mask shape', mask.shape
else:
mask = np.zeros((height*width, number), dtype=np.float32)
maskt = np.reshape(imread('emma/emma_test_mask.jpg').astype(np.float32), (height*width,))
maskt = maskt > 100
for d in xrange(number):
mask[:,d] = maskt
print 'Mask shape', mask.shape
if i == 0:
mask = tf.constant(mask)
print 'Mask shape', map(lambda i: i.value, mask.get_shape())
feats = tf.mul(feats,mask)
gram = tf.matmul(tf.transpose(feats), feats) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
else:
mask2 = mask < 1
feats2 = tf.mul(feats,mask2)
gram2 = tf.matmul(tf.transpose(feats2), feats2) / size
style_gram = style_features[i][style_layer]
style_losses.append(2 * tf.nn.l2_loss(gram2 - style_gram) / style_gram.size)
style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)
# total variation denoising
tv_y_size = _tensor_size(image[:,1:,:,:])
tv_x_size = _tensor_size(image[:,:,1:,:])
tv_loss = tv_weight * 2 * (
(tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
tv_y_size) +
(tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
tv_x_size))
# overall loss
loss = content_loss + style_loss + tv_loss
# optimizer setup
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
def print_progress(i, last=False):
if print_iterations is not None:
if i is not None and i % print_iterations == 0 or last:
print >> stderr, ' content loss: %g' % content_loss.eval()
print >> stderr, ' style loss: %g' % style_loss.eval()
print >> stderr, ' tv loss: %g' % tv_loss.eval()
print >> stderr, ' total loss: %g' % loss.eval()
# optimization
best_loss = float('inf')
best = None
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(iterations):
print_progress(i)
print >> stderr, 'Iteration %d/%d' % (i + 1, iterations)
train_step.run()
if (checkpoint_iterations is not None and
i % checkpoint_iterations == 0) or i == iterations - 1:
this_loss = loss.eval()
if this_loss < best_loss:
best_loss = this_loss
best = image.eval()
print_progress(None, i == iterations - 1)
if i % 10 == 0 and best is not None:
tmp_img = vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
imsave("iter" + str(i) + ".jpg", tmp_img)
return vgg.unprocess(best.reshape(shape[1:]), mean_pixel)
def stylize(
network,
initial,
content,
style,
iterations,
content_weight,
style_weight,
tv_weight,
learning_rate,
print_iter=None,
):
shape = (1,) + content.shape
style_shape = (1,) + style.shape
content_features = {}
style_features = {}
g = tf.Graph()
with g.as_default(), g.device("/cpu:0"), tf.Session() as sess:
image = tf.placeholder("float", shape=shape)
net, mean_pixel = vgg.net(network, image)
content_pre = np.array([vgg.preprocess(content, mean_pixel)])
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval(feed_dict={image: content_pre})
g = tf.Graph()
with g.as_default(), g.device("/cpu:0"), tf.Session() as sess:
image = tf.placeholder("float", shape=style_shape)
net, _ = vgg.net(network, image)
style_pre = np.array([vgg.preprocess(style, mean_pixel)])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={image: style_pre})
features = np.reshape(features, (-1, features.shape[3]))
gram = np.matmul(features.T, features) / (features.size)
style_features[layer] = gram
with tf.Graph().as_default():
if initial is None:
noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
initial = tf.random_normal(shape) * 256 / 1000
else:
initial = np.array([vgg.preprocess(initial, mean_pixel)])
initial = initial.astype("float32")
image = tf.Variable(initial)
net, _ = vgg.net(network, image)
content_loss = tf.nn.l2_loss(net[CONTENT_LAYER] - content_features[CONTENT_LAYER])
style_losses = []
for i in STYLE_LAYERS:
layer = net[i]
_, height, width, number = map(lambda i: i.value, layer.get_shape())
size = height * width * number
feats = tf.reshape(layer, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / (size)
style_gram = style_features[i]
style_losses.append(tf.nn.l2_loss(gram - style_gram))
style_loss = reduce(tf.add, style_losses) / len(style_losses)
tv_loss = tf.nn.l2_loss(image[:, 1:, :, :] - image[:, : shape[1] - 1, :, :]) + tf.nn.l2_loss(
image[:, :, 1:, :] - image[:, :, : shape[2] - 1, :]
)
loss = content_weight * content_loss + style_weight * style_loss + tv_weight * tv_loss
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for i in range(iterations):
if print_iter is not None and i % print_iter == 0:
print " content loss: %g" % (content_loss.eval())
print " style loss: %g" % (style_loss.eval())
print " tv loss: %g" % (tv_loss.eval())
print " total loss: %g" % loss.eval()
print "Iteration %d/%d" % (i + 1, iterations)
train_step.run()
return vgg.unprocess(image.eval().reshape(shape[1:]), mean_pixel)