Question

这是我的功能提取网络。

class vgg16:
    def __init__(self, imgs1,imgs2, weights=None, sess=None):

        self.imgs1 = imgs1
    self.imgs2 = imgs2

        with tf.variable_scope("siamese") as scope:
            self.o1 = self.convlayers(self.imgs1)

        .....
        if weights is not None and sess is not None:
                self.load_weights(weights, sess)
            scope.reuse_variables()
            self.o2 = self.convlayers(self.imgs2)

        .......

        if weights is not None and sess is not None:
             self.load_weights(weights, sess)
        #create loss function
        with tf.variable_scope("loss") as scope:

                self.cd=self.cos_dist(self.o1,self.o2,sess)
                self.loss(self.cd,self.o1,self.o2)

    def convlayers(self,_image):
        .....

    def load_weights(self, weight_file, sess):
       ....
    def loss(self,la,i1,j2):
    label=tf.cast(la, tf.float32)
    i=tf.nn.l2_normalize(i1, 1, epsilon=1e-2, name='normed')
    j=tf.nn.l2_normalize(j2, 1, epsilon=1e-2, name='normed')


    sub= tf.subtract(i,j)
    magi=tf.sqrt(tf.reduce_sum(tf.square(sub)))
    ab=tf.abs(magi)
    sq=tf.square(ab)
    self.left=sq*label
    #working on right now
    second_part=1-label
    tpart=tf.constant(0.675, dtype=tf.float32, name='T')-magi
    tp=tf.maximum(0.0,tpart)
    self.right=second_part*tp
    self.final_loss=self.left+self.right

    o= tf.train.GradientDescentOptimizer(0.09)
        self.l=o.compute_gradients(self.final_loss)
        self.o=o.apply_gradients(self.l)




    def cos_dist(self,net_1,net_2,sess):
    ......


if __name__ == '__main__':
  # for batch processing


    sess = tf.InteractiveSession()
    imgs1 = tf.placeholder(tf.float32, [None, 224, 224, 3])
    imgs2 = tf.placeholder(tf.float32, [None, 224, 224, 3])
    vgg = vgg16(imgs1,imgs2, 'vgg16_weights.npz', sess)
    train_step = tf.train.GradientDescentOptimizer(0.001).minimize(vgg.final_loss)



    #try:
    img1 = imread('c2.jpeg', mode='RGB')

    img1 = imresize(img1,(224, 224))
    img2 = imread('c7.jpeg', mode='RGB')

    img2 = imresize(img2,(224, 224))

    M= sess.run([vgg.k,vgg.o,vgg.final_loss],feed_dict={vgg.imgs1: [img1],vgg.imgs2: [img2]})

这是一个亚洲时尚的CNN。问题是后面的层没有得到更新。第一批更新为Nan。 sess.run（）有什么问题吗？我调用属性的方式。通常其中一个原因是输入错误。我尝试打印值img1和img2，矩阵中包含的值仅为255

在backpropogation期间，网络中的权重不会更新

0 个答案: