张量流中成本函数输出'nan'

Question

在研究张量流时，我遇到了一个问题。
成本函数输出'nan'。

而且，如果您在源代码中发现任何其他错误，请让我知道它的链接。

我试图将成本函数值发送给我的训练模型，但它不工作。

tf.reset_default_graph() 

tf.set_random_seed(777) 

X = tf.placeholder(tf.float32, [None, 20, 20, 3]) 
Y = tf.placeholder(tf.float32, [None, 1]) 

with tf.variable_scope('conv1') as scope: 
    W1 = tf.Variable(tf.random_normal([4, 4, 3, 32], stddev=0.01), name='weight1')  
    L1 = tf.nn.conv2d(X, W1, strides=[1, 1, 1, 1], padding='SAME') 
    L1 = tf.nn.relu(L1) 
    L1 = tf.nn.max_pool(L1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') 
    L1 = tf.reshape(L1, [-1, 10 * 10 * 32]) 

    W1_hist = tf.summary.histogram('conv_weight1', W1) 
    L1_hist = tf.summary.histogram('conv_layer1', L1) 

with tf.name_scope('fully_connected_layer1') as scope: 
    W2 = tf.get_variable('W2', shape=[10 * 10 * 32, 1], initializer=tf.contrib.layers.xavier_initializer())   
    b = tf.Variable(tf.random_normal([1])) 
    hypothesis = tf.matmul(L1, W2) + b 

    W2_hist = tf.summary.histogram('fully_connected_weight1', W2) 
    b_hist = tf.summary.histogram('fully_connected_bias', b) 
    hypothesis_hist = tf.summary.histogram('hypothesis', hypothesis) 

with tf.name_scope('cost') as scope: 
    cost = -tf.reduce_mean(Y * tf.log(hypothesis) + (1 - Y) * tf.log(1 - hypothesis)) 
    cost_summary = tf.summary.scalar('cost', cost) 

with tf.name_scope('train_optimizer') as scope: 
    optimizer = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(cost) 

predicted = tf.cast(hypothesis > 0.5, dtype=tf.float32) 
accuracy = tf.reduce_mean(tf.cast(tf.equal(predicted, Y), dtype=tf.float32)) 
accuracy_summary = tf.summary.scalar('accuracy', accuracy) 

train_data_batch, train_labels_batch = tf.train.batch([train_data, train_labels], enqueue_many=True , batch_size=100, allow_smaller_final_batch=True) 

with tf.Session() as sess: 
    # tensorboard --logdir=./logs/planesnet2_log 
    merged_summary = tf.summary.merge_all() 
    writer = tf.summary.FileWriter('./logs/planesnet2_log') 
    writer.add_graph(sess.graph) 

    sess.run(tf.global_variables_initializer()) 
    coord = tf.train.Coordinator() 
    threads = tf.train.start_queue_runners(coord=coord) 
    total_cost = 0 

    for step in range(20): 
     x_batch, y_batch = sess.run([train_data_batch, train_labels_batch]) 
     feed_dict = {X: x_batch, Y: y_batch} 
     _, cost_val = sess.run([optimizer, cost], feed_dict = feed_dict) 
     total_cost += cost_val 
     print('total_cost: ', total_cost, 'cost_val: ', cost_val) 
    coord.request_stop() 
    coord.join(threads) 

Answer 1

您使用交叉熵损失没有乙状结肠激活功能hypothesis，这样你的价值不是在]界0,1]。日志功能没有为负值定义，它最有可能获得一些。添加sigmoid和epsilon因子，以避免负值或0值，你应该没问题。

Answer 2

据我所知，

交叉熵成本函数假设要预测的假设是随机的值。由于交叉熵使用对数函数和公式。因此，交叉熵损失只能用于随机情况。

因此您必须使用softmax函数来计算hypothesis的概率结果。

W2 = tf.get_variable('W2', shape=[10 * 10 * 32, 1], 
initializer=tf.contrib.layers.xavier_initializer())   
b = tf.Variable(tf.random_normal([1])) 

# hypothesis = tf.matmul(L1, W2) + b 
hypothesis = tf.nn.softmax(tf.add(tf.matmul(L1, W2), b)) 
cost = -tf.reduce_mean(Y * tf.log(hypothesis) + (1 - Y) * tf.log(1 - hypothesis)) 

或者您可以使用此代码

W2 = tf.get_variable('W2', shape=[10 * 10 * 32, 1], 
initializer=tf.contrib.layers.xavier_initializer())   
b = tf.Variable(tf.random_normal([1])) 

hypothesis = tf.matmul(L1, W2) + b 
cost = tf.nn.softmax_cross_entropy_with_logits(labels=Y, logits=hypothesis)