实现的实验代码是这两天下班后顺着思路往下写的，比较丑陋，回头整理封装一下。有几个值得一提的细节： 1 )虽然标准的VGG16图像识别应用中输入图片尺寸被限制在了224x224，但在这里你可以想用多大用多大（因为forward pass不会传到FC layer，btw无法想象生成224x224的艺术图有卵用？）；2）原文使用的L-BFGS在tensorflow中并没有加入，所以使用AdamOptimizer取代；3）由于要对输入图片求导，所以不要用XXXXXnstant作为输入。。4）这玩意儿调参很麻烦
计算content feature的代码：
<code class="language-python">## compute content feature
with tf.Graph().as_default():
image=tf.constant(content,dtype=tf.float32)
with slim.arg_scope(vgg.vgg_arg_scope()):
_, end_points=vgg.vgg_16(image,spatial_squeeze=False)
## load vgg16 model
init_fn = slim.assign_from_checkpoint_fn(
os.path.join('./model/', 'vgg_16.ckpt'),
slim.get_model_variables())
## compute content Feature
content_feature=tf.reshape(end_points['vgg_16/conv4/conv4_2'],shape=[-1])
with tf.Session() as sess:
init_fn(sess)
content_ft_val=sess.run(content_feature)
</code>
计算style feature的代码（用了许多hard-code）：
<code class="language-python">style_list=['vgg_16/conv1/conv1_1',
'vgg_16/conv2/conv2_1',
'vgg_16/conv3/conv3_1',
'vgg_16/conv4/conv4_1',
'vgg_16/conv5/conv5_1']
## compute style feature
def style_gram_compute(idx):
dim=end_points[style_list[idx]].get_shape().as_list()
style_feature=tf.reshape(end_points[style_list[idx]],shape=[-1,dim[-1]])
style_gram=tf.reshape(
tf.matmul(style_feature,style_feature,transpose_a=True),
shape=[-1])/(dim[1]*dim[2])
return style_gram,dim
## compute style feature
with tf.Graph().as_default():
image=tf.constant(style,dtype=tf.float32)
with slim.arg_scope(vgg.vgg_arg_scope()):
_, end_points=vgg.vgg_16(image,spatial_squeeze=False)
## load vgg16 model
init_fn = slim.assign_from_checkpoint_fn(
os.path.join('./model/', 'vgg_16.ckpt'),
slim.get_model_variables())
## compute content Feature
style_gram_1,dim1=style_gram_compute(0)
style_gram_2,dim2=style_gram_compute(1)
style_gram_3,dim3=style_gram_compute(2)
style_gram_4,dim4=style_gram_compute(3)
style_gram_5,dim5=style_gram_compute(4)
with tf.Session() as sess:
init_fn(sess)
style_gram_val_1=sess.run(style_gram_1)
style_gram_val_2=sess.run(style_gram_2)
style_gram_val_3=sess.run(style_gram_3)
style_gram_val_4=sess.run(style_gram_4)
style_gram_val_5=sess.run(style_gram_5)
</code>
主优化程序：
<code class="language-python">def style_loss_compute(style_gram_val,target_style_gram):
style_gram_const=tf.constant(style_gram_val)
style_loss=tf.reduce_mean(tf.squared_difference(target_style_gram,style_gram_const))
return style_loss
with tf.Graph().as_default():
### generate target image ####
image=tf.Variable(tf.random_normal(shape=(1,h,w,3),stddev=np.std(content)*0.1,dtype=tf.float32))
with slim.arg_scope(vgg.vgg_arg_scope()):
_, end_points=vgg.vgg_16(image,spatial_squeeze=False)
## load vgg16 model
init_fn = slim.assign_from_checkpoint_fn(
os.path.join('./model/', 'vgg_16.ckpt'),
slim.get_model_variables())
## compute target content Feature
target_content_feature=tf.reshape(end_points['vgg_16/conv4/conv4_2'],shape=[-1])
## compute target style Feature
target_style_gram_1,_=style_gram_compute(0)
target_style_gram_2,_=style_gram_compute(1)
target_style_gram_3,_=style_gram_compute(2)
target_style_gram_4,_=style_gram_compute(3)
target_style_gram_5,_=style_gram_compute(4)
## define content loss
content_loss=tf.reduce_sum(tf.squared_difference(target_content_feature,content_ft_val))
## define style loss
style_loss_1=style_loss_compute(style_gram_val_1,target_style_gram_1)
style_loss_2=style_loss_compute(style_gram_val_2,target_style_gram_2)
style_loss_3=style_loss_compute(style_gram_val_3,target_style_gram_3)
style_loss_4=style_loss_compute(style_gram_val_4,target_style_gram_4)
style_loss_5=style_loss_compute(style_gram_val_5,target_style_gram_5)
w0=1.0/5
style_loss=w0*style_loss_1+w0*style_loss_2+w0*style_loss_3+w0*style_loss_4+w0*style_loss_5
## define total loss
total_loss=1e-3*content_loss+style_loss
## define train step
global_step = tf.Variable(0, trainable=False)
starter_learning_rate = 1.
learning_rate = tf.train.exponential_decay(starter_learning_rate, global_step,
100, 0.95, staircase=True)
train_op=tf.train.AdamOptimizer(learning_rate).minimize(loss=total_loss,
var_list=[image],
global_step=global_step)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
init_fn(sess)
max_iter=1000
for iter in range(max_iter):
_,loss_content_val,loss_style_val=sess.run([train_op,content_loss,style_loss])
sys.stdout.write('\r'+'iteration:%4d-----Loss_content: %0f6-----Loss_style: %0f6-----learning_rate: %0f5'
%(iter,loss_content_val,loss_style_val,learning_rate.eval()))
output=image.eval()
</code>
200字以内，仅用于支线交流，主线讨论请采用回复功能。