Skip to main content
 首页 » 编程设计

python之对比损失函数适用于孪生网络和优化器有问题

2024年12月31日8虾米姐

我已经构建了一个 Siamese 网络用于练习,我已经检查了网站上的一些示例,但是当我尝试运行代码时它显示了一些错误,我不知道维度是否有问题。希望有人能帮助和教我。

batch_size = 32 
epoch = 50 
kernel_size = (3,3) 
pooling_size = (2,2) 
learning_rate = 0.001 
margin_siamese = 1 
 
 
siamese_nn = tf.Graph() 
with siamese_nn.as_default(): 
 
with tf.name_scope('input'): 
    anchor_img = tf.placeholder(dtype = tf.float32, shape = [None, 28, 28, 1], name = 'anchor_image') 
    compare_img = tf.placeholder(dtype = tf.float32, shape = [None, 28, 28, 1], name = 'compare_image') 
    Y_label = tf.placeholder(dtype = tf.int32, shape = [None, 1], name = 'Y_label') 
    learning_rate = tf.placeholder(dtype = tf.float32, shape = [1]) 
 
with tf.variable_scope('siamese_net'): 
    conv_s1_1 = tf.layers.conv2d(anchor_img, filters= 64, kernel_size= kernel_size, activation= tf.nn.relu, name= 'siamese_1') 
    conv_s1_1 = tf.layers.max_pooling2d(conv_s1_1, pool_size= pooling_size, strides = (1,1), name= 'siamese_1') 
 
    conv_s1_2 = tf.layers.conv2d(conv_s1_1, filters = 32, kernel_size = kernel_size, activation= tf.nn.relu, name= 'siamese_2') 
    conv_s1_2 = tf.layers.max_pooling2d(conv_s1_2, pool_size = pooling_size, strides = (1,1), name= 'siamese_2') 
 
    conv_s1_3 = tf.layers.conv2d(conv_s1_2, filters = 16, kernel_size = kernel_size, activation= tf.nn.relu, name= 'siamese_3') 
    conv_s1_3 = tf.layers.max_pooling2d(conv_s1_3, pool_size = pooling_size, strides = (1,1), name= 'siamese_3') 
 
    anchor_vector = tf.layers.flatten(conv_s1_3, name = 'flatten') 
 
with tf.variable_scope('siamese_net', reuse= True): 
 
    conv_s2_1 = tf.layers.conv2d(compare_img, filters= 64, kernel_size= kernel_size, activation= tf.nn.relu, name= 'siamese_1') 
    conv_s2_1 = tf.layers.max_pooling2d(conv_s2_1, pool_size= pooling_size, strides = (1,1), name= 'siamese_1') 
 
    conv_s2_2 = tf.layers.conv2d(conv_s2_1, filters = 32, kernel_size = kernel_size, activation= tf.nn.relu, name= 'siamese_2') 
    conv_s2_2 = tf.layers.max_pooling2d(conv_s2_2, pool_size = pooling_size, strides = (1,1), name= 'siamese_2') 
 
    conv_s2_3 = tf.layers.conv2d(conv_s2_2, filters = 16, kernel_size = kernel_size, activation= tf.nn.relu, name= 'siamese_3') 
    conv_s2_3 = tf.layers.max_pooling2d(conv_s2_3, pool_size = pooling_size, strides = (1,1), name= 'siamese_3') 
 
    compare_vector = tf.layers.flatten(conv_s2_3, name = 'flatten') 
 
 
with tf.name_scope('loss_function'): 
    loss = tf.contrib.losses.metric_learning.contrastive_loss(Y_label, anchor_vector, compare_vector) 
 
with tf.name_scope('training'): 
    optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate) 
    optimizer.minimize(loss) 

错误信息

Shape must be rank 0 but is rank 1 for 'training/Adam/update_siamese_net/siamese_1/kernel/ApplyAdam' (op: 'ApplyAdam') with input shapes: [3,3,1,64], [3,3,1,64], [3,3,1,64], [], [], [1], [], [], [], [3,3,1,64]. 

请您参考如下方法:

为了社区的利益,在下面提及 giser_yugang 的回答。

学习率被声明为一维张量,而它应该是一个标量。

所以,改变代码行,

learning_rate = tf.placeholder(dtype = tf.float32, shape = [1]) 到

learning_rate = tf.placeholder(dtype = tf.float32, shape = []) 将解决错误。