Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
0 parents
commit 4f1ea8e
Showing
5 changed files
with
228 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,14 @@ | ||
root = true | ||
|
||
[*] | ||
charset = utf-8 | ||
end_of_line = lf | ||
insert_final_newline = true | ||
indent_style = space | ||
trim_trailing_whitespace = true | ||
|
||
[*.py] | ||
indent_size = 4 | ||
|
||
[*.md] | ||
trim_trailing_whitespace = false |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,2 @@ | ||
*.mat | ||
env/ |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,13 @@ | ||
# neural-style | ||
|
||
An implementation of [this paper](http://arxiv.org/pdf/1508.06576v2.pdf) in | ||
TensorFlow. | ||
|
||
## Requirements | ||
|
||
* TensorFlow | ||
* SciPy | ||
* Pillow | ||
* NumPy | ||
* [Pre-trained VGG | ||
network](http://www.vlfeat.org/matconvnet/models/imagenet-vgg-verydeep-19.mat) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,125 @@ | ||
import vgg | ||
|
||
import tensorflow as tf | ||
import numpy as np | ||
import scipy.misc as sm | ||
|
||
import sys | ||
import math | ||
|
||
VGG_PATH = 'imagenet-vgg-verydeep-19.mat' | ||
CONTENT_LAYER = 'relu4_2' | ||
STYLE_LAYERS = ('relu1_1', 'relu2_1', 'relu3_1', 'relu4_1', 'relu5_1') | ||
NOISE_RATIO = 0.0 | ||
ALPHA = 1.0 # weight of content loss | ||
BETA = 1e4 # weight of style loss | ||
LEARNING_RATE_INITIAL = 2e1 | ||
LEARNING_DECAY_BASE = 0.94 | ||
LEARNING_DECAY_STEPS = 100 | ||
|
||
def imread(path): | ||
return sm.imread(path).astype(np.float) | ||
|
||
def imsave(path, img): | ||
img = np.clip(img, 0, 255).astype(np.uint8) | ||
sm.imsave(path, img) | ||
|
||
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: | ||
width = content_image.shape[1] | ||
|
||
content_aspect = (float(content_image.shape[0]) / | ||
content_image.shape[1]) # height / width | ||
new_shape = (int(math.floor(float(content_image.shape[0]) / | ||
content_image.shape[1] * width)), width) | ||
content_image = sm.imresize(content_image, new_shape) | ||
style_aspect = (float(style_image.shape[0]) / | ||
style_image.shape[1]) | ||
if style_scale > 0: | ||
style_image_scaled = sm.imresize(style_image, style_scale) | ||
shape = style_image_scaled.shape | ||
if shape[0] >= new_shape[0] and shape[1] >= new_shape[1]: | ||
style_image = style_image_scaled | ||
else: | ||
style_scale = -1 | ||
if style_scale <= 0: | ||
matched_height = int(math.ceil(new_shape[1] * style_aspect)) | ||
if matched_height >= new_shape[0]: | ||
style_image = sm.imresize(style_image, (matched_height, new_shape[1])) | ||
else: | ||
matched_width = int(math.ceil(new_shape[0] / style_aspect)) | ||
style_image = sm.imresize(style_image, (new_shape[0], matched_width)) | ||
style_image = style_image[0:new_shape[0], 0:new_shape[1], :] | ||
assert content_image.shape == style_image.shape | ||
|
||
shape = (1,) + content_image.shape | ||
|
||
content_features = {} | ||
style_features = {} | ||
g = tf.Graph() | ||
with g.as_default(): | ||
image = tf.placeholder('float', shape=shape) | ||
net, mean_pixel = vgg.net(VGG_PATH, image) | ||
|
||
with tf.Session() as sess: | ||
content_pre = np.array([vgg.preprocess(content_image, mean_pixel)]) | ||
content_features[CONTENT_LAYER] = net[CONTENT_LAYER].eval( | ||
feed_dict={image: content_pre}) | ||
|
||
style_pre = np.array([vgg.preprocess(style_image, mean_pixel)]) | ||
for layer in STYLE_LAYERS: | ||
style_features[layer] = net[layer].eval( | ||
feed_dict={image: style_pre}) | ||
|
||
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) | ||
|
||
match = style_features[i] | ||
match_feats = np.reshape(match, (-1, match.shape[3])) | ||
match_gram = np.matmul(match_feats.T, match_feats) | ||
|
||
style_losses.append(tf.nn.l2_loss(gram - match_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(10000): | ||
print 'i = %d' % i | ||
imsave('%05d.jpg' % i, vgg.unprocess( | ||
image.eval().reshape(shape[1:]), mean_pixel)) | ||
train_step.run() | ||
|
||
|
||
if __name__ == '__main__': | ||
main() |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,74 @@ | ||
import tensorflow as tf | ||
import numpy as np | ||
import scipy.io as sio | ||
|
||
def _conv_layer(weights, bias): | ||
def _make_layer(input): | ||
conv = tf.nn.conv2d(input, tf.constant(weights), strides=[1, 1, 1, 1], | ||
padding='SAME') | ||
return tf.nn.bias_add(conv, bias) | ||
return _make_layer | ||
|
||
def _pool_layer(): | ||
def _make_layer(input): | ||
return tf.nn.max_pool(input, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') | ||
return _make_layer | ||
|
||
def _add_layer(input_image, layers, func): | ||
if not layers: | ||
new = func(input_image) | ||
else: | ||
new = func(layers[-1]) | ||
layers.append(new) | ||
|
||
def preprocess(image, mean_pixel): | ||
return image - mean_pixel | ||
|
||
def unprocess(image, mean_pixel): | ||
image = image + mean_pixel | ||
return image | ||
|
||
|
||
def net(data_path, input_image): | ||
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' | ||
] | ||
|
||
|
||
data = sio.loadmat(data_path) | ||
mean = data['normalization'][0][0][0] | ||
mean_pixel = np.mean(mean, axis=(0, 1)) | ||
constants = data['layers'][0] | ||
|
||
net = [] | ||
for i, kind in enumerate(layers): | ||
short = kind[:4] | ||
if short == 'conv': | ||
weights = constants[i][0][0][0][0][0] | ||
# in matconvnet, weights are [width, height, depth, num_filters] | ||
# but in tensorflow, [height, width, in_channels, out_channels] | ||
weights = np.transpose(weights, (1, 0, 2, 3)) | ||
bias = constants[i][0][0][0][0][1].reshape(-1) | ||
new = _conv_layer(weights, bias) | ||
elif short == 'relu': | ||
new = tf.nn.relu | ||
elif short == 'pool': | ||
new = _pool_layer() | ||
else: | ||
raise ValueError('invalid layer type: %s' % kind) | ||
_add_layer(input_image, net, new) | ||
|
||
assert len(layers) == len(net) | ||
|
||
return dict(zip(layers, net)), mean_pixel |