def create_content_features(content_image):
content_features = {}
with tf.Graph().as_default(), tf.Session() as sess:
X_content = tf.placeholder(tf.float32,
shape=batch_shape,
name="X_content")
X_pre = img_utils.preprocess(X_content)
content_np = np.array([content_image])
content_net, _ = load_graph(
data_path='vgg_weights/imagenet-vgg-verydeep-19.mat',
input_image=X_pre)
content_features[CONTENT_LAYER] = content_net[CONTENT_LAYER].eval(
feed_dict={X_content: content_np})
return content_features
def create_style_features(style_image):
style_features = {}
with tf.Graph().as_default(), tf.Session() as sess:
X_style = tf.placeholder(tf.float32,
shape=batch_shape,
name='style_image')
X_pre = img_utils.preprocess(X_style)
net, _ = load_graph(
data_path='vgg_weights/imagenet-vgg-verydeep-19.mat',
input_image=X_pre)
style_np = np.array([style_image])
for layer in STYLE_LAYERS:
features = net[layer].eval(feed_dict={X_style: style_np})
features = features.reshape((-1, features.shape[3]))
gram = np.matmul(features.T, features) / features.size
style_features[layer] = gram
return style_features
input_image = tf.get_variable(name="image",
trainable=True,
dtype=tf.float32,
shape=batch_shape,
initializer=tf.constant_initializer(
value=0., dtype=tf.float32))
net, _ = load_graph(data_path='vgg_weights/imagenet-vgg-verydeep-19.mat',
input_image=input_image)
l_content = create_content_loss(content_features, net)
l_style = create_style_loss(style_features, net)
total_loss = l_content + 512 * l_style
import matplotlib.pyplot as plt
content_image_file = img_utils.preprocess(content_image_file)
initial_image = img_utils.generate_noise_image(content_image_file,
IMAGE_HEIGHT, IMAGE_WIDTH,
0.1)
#initial_image = img_utils.postprocess(initial_image)#undo the preprocessing;
assign_op = input_image.assign(initial_image)
opt = tf.train.AdamOptimizer(learning_rate=LR).minimize(total_loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
an_image = sess.run(assign_op)
plt.figure()
plt.subplot(121)
plt.imshow(