TensorFlow训练MNIST手写数字识别网络报错：无法将dict转为Tensor

首先，直接点出你报错的核心原因：在构建网络层时，你错误地把存储权重和偏置的字典（比如H1、H2、H3）传给了tf.matmul，但这个函数需要的是经过前一层计算后的激活张量（也就是FinalH1、FinalH2、FinalH3），这就导致TensorFlow无法把字典转换成张量，抛出了类型错误。

除此之外，你的代码还有几个小细节需要修正，我来一步步帮你解决：

1. 修正网络层的计算逻辑 #

在neuralNetwork函数里，每一层的输入必须是前一层输出的激活张量，而不是权重字典：

• 计算FinalH2时，要用FinalH1代替H1
• 计算FinalH3时，要用FinalH2代替H2
• 计算FinalOut时，要用FinalH3代替H3

2. 修正Session的初始化语法 #

with tf.Session as sess:少了括号，正确写法是with tf.Session() as sess:

3. 修正优化器变量名的拼写错误 #

你定义的优化器变量是optimizingTool，但后面调用时写成了optimizer，要保持变量名一致。

修正后的完整代码 #

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets("/tmp/data", one_hot=True)

numNodesH1 = 600
numNodesH2 = 500
numNodesH3 = 500
numNodesOut = 10
sizeOfBatch = 150

y = tf.placeholder("float")
x = tf.placeholder("float", [None, 784])

def neuralNetwork(value):
    # 定义各层的权重和偏置字典
    H1 = {'weights': tf.Variable(tf.random_normal([784, numNodesH1])), 
          "biases": tf.Variable(tf.random_normal([numNodesH1]))}
    H2 = {'weights': tf.Variable(tf.random_normal([numNodesH1, numNodesH2])), 
          "biases": tf.Variable(tf.random_normal([numNodesH2]))}
    H3 = {"weights": tf.Variable(tf.random_normal([numNodesH2, numNodesH3])), 
          "biases": tf.Variable(tf.random_normal([numNodesH3]))}
    output = {"weights": tf.Variable(tf.random_normal([numNodesH3, numNodesOut])), 
              "biases": tf.Variable(tf.random_normal([numNodesOut]))}
    
    # 修正每一层的输入：使用前一层的激活张量而非权重字典
    FinalH1 = tf.add(tf.matmul(value, H1["weights"]), H1["biases"])
    FinalH1 = tf.nn.relu(FinalH1)
    
    FinalH2 = tf.add(tf.matmul(FinalH1, H2["weights"]), H2["biases"])
    FinalH2 = tf.nn.relu(FinalH2)
    
    FinalH3 = tf.add(tf.matmul(FinalH2, H3["weights"]), H3["biases"])
    FinalH3 = tf.nn.relu(FinalH3)
    
    FinalOut = tf.matmul(FinalH3, output["weights"]) + output["biases"]
    return FinalOut

def train(inputdata):
    prediction = neuralNetwork(inputdata)
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y))
    optimizingTool = tf.train.AdamOptimizer().minimize(cost)
    epochsNum = 10
    
    # 修正Session初始化语法
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        for i in range(epochsNum):
            lostEpochs = 0
            for o in range(int(mnist.train.num_examples / sizeOfBatch)):
                ex, ey = mnist.train.next_batch(sizeOfBatch)
                # 修正优化器变量名
                _, c = sess.run([optimizingTool, cost], feed_dict={x: ex, y: ey})
                lostEpochs += c
            print(f"Epochs completed = {i} / {epochsNum}, epoch loss = {lostEpochs}")
        
        correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))
        neuralAccuracy = tf.reduce_mean(tf.cast(correct, "float"))
        print(f"Test Accuracy: {neuralAccuracy.eval({x: mnist.test.images, y: mnist.test.labels})}")

train(x)

额外小提示 #

作为初学者，建议你在构建每一层后，打印张量的形状（比如print(FinalH1.get_shape())），这样能快速发现输入输出维度不匹配或者传错参数的问题，帮你更快排查错误。

内容的提问来源于stack exchange，提问作者LeoTheDev