这个 TensorFlow 代码哪里出错了？

Question

所以我为 CNN 编写了一个代码，图片尺寸为 (350, 1, 1000, 1)，采用 NHWC 格式。现在数据很完美，网络正在运行，但输出很奇怪。这是我的代码（可能没有语法错误，但有一些逻辑错误）：

weights_conv1 = tf.get_variable(name = 'wc1', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 1, 20), mean = 0, stddev = 1.0))
weights_conv2 = tf.get_variable(name = 'wc2', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 20), mean = 0, stddev = 1.0))
weights_conv3 = tf.get_variable(name = 'wc3', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 20), mean = 0, stddev = 1.0))
weights_conv4 = tf.get_variable(name = 'wc4', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 20), mean = 0, stddev = 1.0))
weights_conv5 = tf.get_variable(name = 'wc5', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 10), mean = 0, stddev = 1.0))

filters = [weights_conv1] + [weights_conv2] + [weights_conv3] + [weights_conv4] + [weights_conv5]

bias1 = tf.get_variable(name = 'b1', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias2 = tf.get_variable(name = 'b2', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias3 = tf.get_variable(name = 'b3', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias4 = tf.get_variable(name = 'b4', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias5 = tf.get_variable(name = 'b5', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 10)))

biases = [bias1] + [bias2] + [bias3] + [bias4] + [bias5]


def convolutionForwardPropagation(img, filters, biases):
    c1 = tf.nn.conv2d(img, filters[0], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f1 = tf.nn.relu(c1) + biases[0]
    c2 = tf.nn.conv2d(f1, filters[1], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f2 = tf.nn.relu(c2) + biases[1]
    
    c3 = tf.nn.conv2d(f2, filters[2], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f3 = tf.nn.relu(c3) + biases[2]
    c4 = tf.nn.conv2d(f3, filters[3], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f4 = tf.nn.relu(c4) + biases[3]
        
    c5 = tf.nn.conv2d(f4, filters[4], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f5 = tf.nn.relu(c5) + biases[4]
    
    shape = f5.shape
    fr = tf.reshape(f5,(shape[0], shape[3] * shape[2]))
    fc1 = tf.contrib.layers.fully_connected(fr, 1000, activation_fn = tf.nn.relu)
    fc2 = tf.contrib.layers.fully_connected(fc1, 2, activation_fn = tf.nn.relu)
    print(fc2.shape)
    
    return fc2

fc = convolutionForwardPropagation(trainDataset, filters, biases)
entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits = fc, labels = labelsTrain, name = 'cross_entropy')
loss = tf.reduce_mean(entropy, name = 'loss')

optimizer = tf.train.AdamOptimizer(learning_rate).minimize(loss)

hypothesis = tf.nn.softmax(fc)
correct_preds = tf.equal(tf.argmax(hypothesis, 1), tf.argmax(labelsTrain, 1))
accuracy = tf.reduce_sum(tf.cast(correct_preds, tf.float32))


writer = tf.summary.FileWriter('./graphs/logreg', tf.get_default_graph())
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(0, epochs):
        sess.run(fc)
        _, l = sess.run([optimizer, loss])                 
        sess.run(hypothesis)
        sess.run(correct_preds)
        acc = sess.run(accuracy)
        print("Epoch :", i+1, ", loss : ", l, ", accuracy :", acc)
writer.close()

结果有点奇怪：

(350, 2)
Epoch : 1 , loss :  599.7743 , accuracy : 175.0
Epoch : 2 , loss :  64824.633 , accuracy : 175.0
Epoch : 3 , loss :  15540.435 , accuracy : 175.0
Epoch : 4 , loss :  0.69314754 , accuracy : 175.0
Epoch : 5 , loss :  0.69314754 , accuracy : 175.0
Epoch : 6 , loss :  0.69314754 , accuracy : 175.0
Epoch : 7 , loss :  0.69314754 , accuracy : 175.0
Epoch : 8 , loss :  0.69314754 , accuracy : 175.0
.....

那么我的代码到底出了什么问题？

Answer 1

您需要为输入数据和标签创建占位符。

然后您需要使用这些占位符作为函数的输入，并在 sess.run() 中为这些占位符输入值。

以下是您的修复代码。

eights_conv1 = tf.get_variable(name = 'wc1', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 1, 20), mean = 0, stddev = 1.0))
weights_conv2 = tf.get_variable(name = 'wc2', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 20), mean = 0, stddev = 1.0))
weights_conv3 = tf.get_variable(name = 'wc3', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 20), mean = 0, stddev = 1.0))
weights_conv4 = tf.get_variable(name = 'wc4', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 20), mean = 0, stddev = 1.0))
weights_conv5 = tf.get_variable(name = 'wc5', dtype = tf.float32, initializer = tf.random_normal(shape = (1, 3, 20, 10), mean = 0, stddev = 1.0))

filters = [weights_conv1] + [weights_conv2] + [weights_conv3] + [weights_conv4] + [weights_conv5]

bias1 = tf.get_variable(name = 'b1', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias2 = tf.get_variable(name = 'b2', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias3 = tf.get_variable(name = 'b3', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias4 = tf.get_variable(name = 'b4', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 20)))
bias5 = tf.get_variable(name = 'b5', dtype = tf.float32, initializer = tf.random_normal(mean = 0, stddev = 1.0, shape = (1, 1, 1, 10)))

biases = [bias1] + [bias2] + [bias3] + [bias4] + [bias5]

X = tf.placeholder(tf.float32, shape=[350, 1, 1000, 1],name = "img_trainDataset")
Y = tf.placeholder(tf.float32, shape=[dim1,dim2],name = "labelsTrain")

def convolutionForwardPropagation(X, filters, biases):
    c1 = tf.nn.conv2d(X, filters[0], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f1 = tf.nn.relu(c1) + biases[0]
    c2 = tf.nn.conv2d(f1, filters[1], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f2 = tf.nn.relu(c2) + biases[1]

    c3 = tf.nn.conv2d(f2, filters[2], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f3 = tf.nn.relu(c3) + biases[2]
    c4 = tf.nn.conv2d(f3, filters[3], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f4 = tf.nn.relu(c4) + biases[3]

    c5 = tf.nn.conv2d(f4, filters[4], strides =[1,1,1,1], data_format ='NHWC', padding = 'VALID')
    f5 = tf.nn.relu(c5) + biases[4]

    shape = f5.shape
    fr = tf.reshape(f5,(shape[0], shape[3] * shape[2]))
    fc1 = tf.contrib.layers.fully_connected(fr, 1000, activation_fn = tf.nn.relu)
    fc2 = tf.contrib.layers.fully_connected(fc1, 2, activation_fn = tf.nn.relu)
    print(fc2.shape)

    return fc2

fc = convolutionForwardPropagation(X, filters, biases)
entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits = fc, labels = Y, name = 'cross_entropy')
loss = tf.reduce_mean(entropy, name = 'loss')

optimizer = tf.train.AdamOptimizer(learning_rate).minimize(loss)

hypothesis = tf.nn.softmax(fc)
correct_preds = tf.equal(tf.argmax(hypothesis, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_sum(tf.cast(correct_preds, tf.float32))


writer = tf.summary.FileWriter('./graphs/logreg', tf.get_default_graph())
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(0, epochs):
        _, l, h, c, a = sess.run([optimizer, loss, hypothesis, correct_preds, accuracy], feed_dict={X: trainDataset, Y: labelsTrain})                 
        print("Epoch :", i+1, ", loss : ", l, ", accuracy :", a)
writer.close()

作为一般经验法则，尽量不要使用多个 sess.run()，而是将您需要的所有内容放在一个 sess.run() 中。每次调用 sess.run() 都会执行整个图形，因此计算量很大。