下面有两个代码段来自仓库https://github.com/kang205/DVBPR。这是本文的python实现：
具有生成图像模型的视觉感知时尚推荐和设计
我正在尝试运行，理解并调整它。据我在论文中了解的那样，他们首先训练了一个暹罗CNN网络（DVBPR main.py代码）作为偏好预测器。然后是经过训练的gan网络，可以生成与客户喜欢的图像相似的图像，最后，偏好最大化块（pm main.py）运行正则化过程，以防止使用偏好预测变量的矩阵分解模型过拟合。
我试图跳过gan部分，只是从我之前训练的暹罗网络中获取新图像的偏好得分。这就是我在下面的“优先代码”部分中要执行的操作。我从经过训练的暹罗网络重新加载检查点权重，声明一些变量，最后将用户未看到的图像和idx传递给会话。
我不完全了解idx的用途或目的。如果有人能解释我将不胜感激。
我也在尝试了解在打印sess.run时输出中返回的内容。是我通过sess.run和首选项得分获得的图像的图像张量吗？我注意到是否更改了sess.run中括号中的顺序，它会更改输出的顺序。因此，数组（[99.19335938]，dtype = float32）从idx开始。
我是tensorFlow的新手，因此非常感谢所有提示。
暹罗网络代码（DVBPR main.py）：
import sys
import math
import random
import time
from PIL import Image
import Queue
import numpy as np
import threading
from cStringIO import StringIO
import tensorflow as tf
dataset_name = 'AmazonFashion6ImgPartitioned.npy'
#Hyper-prameters
K = 100 # Latent dimensionality
lambda1 = 0.001 # Weight decay
lambda2 = 1.0 # Regularizer for theta_u
learning_rate = 1e-4
training_epoch = 20
batch_size = 128
dropout = 0.5 # Dropout, probability to keep units
numldprocess=4 # multi-threading for loading images
dataset = np.load('../'+dataset_name)
[user_train, user_validation, user_test, Item, usernum, itemnum] = dataset
# Create some wrappers for simplicity
def conv2d(x, W, b, strides=1):
# Conv2D wrapper, with bias and relu activation
x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
x = tf.nn.bias_add(x, b)
return tf.nn.relu(x)
def maxpool2d(x, k=2):
return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
padding='SAME')
def avgpool2d(x, k=2):
return tf.nn.avg_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
padding='SAME')
weights = {
'wc1': [11, 11, 3, 64],
'wc2': [5, 5, 64, 256],
'wc3': [3, 3, 256, 256],
'wc4': [3, 3, 256, 256],
'wc5': [3, 3, 256, 256],
'wd1': [7*7*256, 4096],
'wd2': [4096, 4096],
'wd3': [4096, K],
}
biases = {
'bc1': [64],
'bc2': [256],
'bc3': [256],
'bc4': [256],
'bc5': [256],
'bd1': [4096],
'bd2': [4096],
'bd3': [K],
}
def Weights(name):
return tf.get_variable(name,dtype=tf.float32,shape=weights[name],initializer=tf.contrib.layers.xavier_initializer())
def Biases(name):
return tf.get_variable(name,dtype=tf.float32,initializer=tf.zeros(biases[name]))
# Create CNN model
def CNN(x,dropout):
# Reshape input picture
x = tf.reshape(x, shape=[-1, 224, 224, 3])
conv1 = conv2d(x, Weights('wc1'), Biases('bc1'), strides=4)
conv1 = tf.nn.relu(conv1)
conv1 = maxpool2d(conv1, k=2)
conv2 = conv2d(conv1, Weights('wc2'), Biases('bc2'))
conv2 = tf.nn.relu(conv2)
conv2 = maxpool2d(conv2, k=2)
conv3 = conv2d(conv2, Weights('wc3'), Biases('bc3'))
conv3 = tf.nn.relu(conv3)
conv4 = conv2d(conv3, Weights('wc4'), Biases('bc4'))
conv4 = tf.nn.relu(conv4)
conv5 = conv2d(conv4, Weights('wc5'), Biases('bc5'))
conv5 = tf.nn.relu(conv5)
conv5 = maxpool2d(conv5, k=2)
fc1 = tf.reshape(conv5, [-1,weights['wd1'][0]])
fc1 = tf.add(tf.matmul(fc1, Weights('wd1')), Biases('bd1'))
fc1 = tf.nn.relu(fc1)
fc1 = tf.nn.dropout(fc1, dropout)
fc2 = tf.add(tf.matmul(fc1, Weights('wd2')), Biases('bd2'))
fc2 = tf.nn.relu(fc2)
fc2 = tf.nn.dropout(fc2, dropout)
fc3 = tf.add(tf.matmul(fc2, Weights('wd3')), Biases('bd3'))
return fc3
image1=(tf.to_float(image1)-127.5)/127.5
image2=(tf.to_float(image2)-127.5)/127.5
_image_test=(tf.to_float(image_test)-127.5)/127.5
u=tf.reshape(u,shape=[batch_size])
i=tf.reshape(i,shape=[batch_size])
j=tf.reshape(j,shape=[batch_size])
keep_prob = tf.placeholder(tf.float32) #dropout (keep probability)
#siamese networks
with tf.variable_scope("DVBPR") as scope:
result1 = CNN(image1,dropout)
scope.reuse_variables()
result2 = CNN(image2,dropout)
result_test = CNN(_image_test,1.0)
nn_regularizers = sum(map(tf.nn.l2_loss,[Weights('wd1'), Weights('wd2'), Weights('wd3'), Weights('wc1'), Weights('wc2'), Weights('wc3'), Weights('wc4'), Weights('wc5')]))
thetau = tf.Variable(tf.random_uniform([usernum,K],minval=0,maxval=1)/100)
cost_train = tf.reduce_sum(tf.log(tf.sigmoid(tf.reduce_sum(tf.multiply(tf.gather(thetau,u),tf.subtract(result1,result2)),1,keep_dims=True))))
regularizers = tf.nn.l2_loss(tf.gather(thetau,u))
cost_train -= lambda1 * nn_regularizers + lambda2 * regularizers
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(-cost_train)
# Initializing the variables
init = tf.initialize_all_variables()
def AUC(train,test,U,I):
ans=0
cc=0
for u in train:
i=test[u][0]['productid']
T=np.dot(U[u,:],I.T)
cc+=1
M=set()
for item in train[u]:
M.add(item['productid'])
M.add(i)
count=0
tmpans=0
#for j in xrange(itemnum):
for j in random.sample(xrange(itemnum),100): #sample
if j in M: continue
if T[i]>T[j]: tmpans+=1
count+=1
tmpans/=float(count)
ans+=tmpans
ans/=float(cc)
return ans
def Evaluation(step):
print '...'
U=sess.run(thetau)
I=np.zeros([itemnum,K],dtype=np.float32)
idx=np.array_split(range(itemnum),(itemnum+batch_size-1)/batch_size)
input_images=np.zeros([batch_size,224,224,3],dtype=np.int8)
for i in range(len(idx)):
cc=0
for j in idx[i]:
input_images[cc]=np.uint8(np.asarray(Image.open(StringIO(Item[j]['imgs'])).convert('RGB').resize((224,224))))
cc+=1
I[idx[i][0]:(idx[i][-1]+1)]=sess.run(result_test,feed_dict={image_test:input_images})[:(idx[i][-1]-idx[i][0]+1)]
print 'export finised!'
np.save('UI_'+str(K)+'_'+str(step)+'.npy',[U,I])
return AUC(user_train,user_validation,U,I), AUC(user_train,user_test,U,I)
def sample(user):
u = random.randrange(usernum)
numu = len(user[u])
i = user[u][random.randrange(numu)]['productid']
M=set()
for item in user[u]:
M.add(item['productid'])
while True:
j=random.randrange(itemnum)
if (not j in M): break
return (u,i,j)
def load_image_async():
while True:
(uuu,iii,jjj)=sample(user_train)
jpg1=np.uint8(np.asarray(Image.open(StringIO(Item[iii]['imgs'])).convert('RGB').resize((224,224))))
jpg2=np.uint8(np.asarray(Image.open(StringIO(Item[jjj]['imgs'])).convert('RGB').resize((224,224))))
sess.run(batch_train_queue_op,feed_dict={queueu:np.asarray([uuu]),
queuei:np.asarray([iii]),
queuej:np.asarray([jjj]),
queueimage1:jpg1,queueimage2:jpg2,
})
f=open('DVBPR.log','w')
config = tf.ConfigProto(log_device_placement=False,allow_soft_placement=True)
sess=tf.Session(config=config)
sess.run(init)
t=[0]*numldprocess
for i in range(numldprocess):
t[i] = threading.Thread(target=load_image_async)
t[i].daemon=True
t[i].start()
oneiteration = 0
for item in user_train: oneiteration+=len(user_train[item])
step = 1
saver = tf.train.Saver([k for k in tf.global_variables() if k.name.startswith('DVBPR')])
epoch=0
while step * batch_size <= training_epoch*oneiteration+1:
sess.run(optimizer, feed_dict={keep_prob: dropout})
print 'Step#'+str(step)+' CNN update'
if step*batch_size / oneiteration >epoch:
epoch+=1
saver.save(sess,'./DVBPR_auc_'+str(K)+'_'+str(step)+'.ckpt')
auc_valid,auc_test=Evaluation(step)
print 'Epoch #'+str(epoch)+':'+str(auc_test)+' '+str(auc_valid)+'\n'
f.write('Epoch #'+str(epoch)+':'+str(auc_test)+' '+str(auc_valid)+'\n')
f.flush()
step += 1
print "Optimization Finished!"
偏好得分的代码（pm main.py）：
from cStringIO import StringIO
def loadimg(item):
return np.round(np.array(Image.open(StringIO(Item[item]['imgs'])).convert('RGB').resize((224,224)),dtype=np.float64))
with tf.device('/cpu:0'):
# creating placeholder variable
# _image_test has form <tf.Tensor 'div_4:0' shape=(64, 224, 224, 3) dtype=float32>
image_test=tf.placeholder(dtype=tf.uint8,shape=[64,224,224,3])
_image_test=(tf.to_float(image_test)-127.5)/127.5
keep_prob = tf.placeholder(tf.float32) #dropout (keep probability)
# sharing variable by capturing scope and setting reuse
with tf.variable_scope("DVBPR") as scope:
# creating model from input image and drop out, returns tensor shape (100,1)
result_test = CNN(_image_test,1.0)
# creating variable thetau which is tensor of shape [usernum,D] filled with random values in range 0 to 1
thetau = tf.Variable(tf.random_uniform([usernum,D],minval=0,maxval=1)/100)
# Initializing the variables
init = tf.initialize_all_variables()
config = tf.ConfigProto(log_device_placement=False,allow_soft_placement=True)
config.gpu_options.allow_growth = True
# creating session
sess=tf.Session(config=config)
sess.run(init)
# creates object to save all variables in
saver = tf.train.Saver([k for k in tf.global_variables() if k.name.startswith('DVBPR')])
# restore model weights from DVBPR check point
saver.restore(sess,DVBPR_ckpt_path)
# creating tensor for image with same shape as gan_image
tstImg3=np.round(np.array(Image.open(StringIO(Item[119]['imgs'])).convert('RGB').resize((224,224)),dtype=np.float32))
rep_gan=tf.reshape(tstImg3, shape=[-1, 224, 224, 3])
# Note: don't have gpu switch to cpu
with tf.device('/cpu:0'):
# need to figure out how to substitute image of product customer hasn't seen for get_gen
gan_image=rep_gan
# shaping image tensor from gan_image
image=tf.image.resize_nearest_neighbor(images=gan_image, size=[224,224], align_corners=None, name=None)
with tf.variable_scope("DVBPR") as scope:
scope.reuse_variables()
# running model defined previously to get tensor from "image" variable above
result = CNN(image,1.0)
_user=12
# idx and user defined in opt section
user=tf.placeholder(dtype=tf.int32,shape=[1])
idx=tf.reduce_sum(tf.matmul(result,tf.transpose(tf.gather(thetau,user))),1)
# Note: I'm not sure what the purpose of idx is
# run piece from Preference Score with DVBPR instead of opt
print(sess.run([gan_image,idx],feed_dict={user:[_user]}))
# the results below appear to return an image tensor and a preference score
Output:
[array([[[[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.],
...,
[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.]],
[[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.],
...,
[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.]],
[[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.],
...,
[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.]],
...,
[[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.],
...,
[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.]],
[[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.],
...,
[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.]],
[[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.],
...,
[ 255., 255., 255.],
[ 255., 255., 255.],
[ 255., 255., 255.]]]], dtype=float32), array([ 99.19335938], dtype=float32)]