python - 无法从 mnist 数据集生成准确的结果

Question

我一直在练习机器学习，并且遇到了 mnist 教程。在学习的过程中，我编写了这段代码。

` 将 tensorflow 导入为 tf 从tensorflow.examples.tutorials.mnist导入input_data 将 numpy 导入为 np

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

n_hidden_layer_1 = 500
n_hidden_layer_2 = 500
n_hidden_layer_3 = 500

n_classes = 10
batch_size = 100

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

hidden_layer_1 = {
    'weights': tf.Variable(tf.random_normal(shape = [784, n_hidden_layer_1])),
    'bias': tf.Variable(tf.random_normal(shape = [n_hidden_layer_1]))
}

hidden_layer_2 = {
    'weights': tf.Variable(tf.random_normal(shape = [n_hidden_layer_1, n_hidden_layer_2])),
    'bias': tf.Variable(tf.random_normal(shape = [n_hidden_layer_2]))
}

hidden_layer_3 = {
    'weights': tf.Variable(tf.random_normal(shape = [n_hidden_layer_2, n_hidden_layer_3])),
    'bias': tf.Variable(tf.random_normal(shape = [n_hidden_layer_3]))
}

output_layer = {
    'weights': tf.Variable(tf.random_normal(shape = [n_hidden_layer_3, n_classes])),
    'bias': tf.Variable(tf.random_normal(shape = [n_classes]))
}

hidden_layer_1_output = tf.nn.relu(tf.add(tf.matmul(x, hidden_layer_1['weights']), hidden_layer_1['bias']))
hidden_layer_2_output = tf.nn.relu(tf.add(tf.matmul(hidden_layer_1_output, hidden_layer_2['weights']), hidden_layer_2['bias']))
hidden_layer_3_output = tf.nn.relu(tf.add(tf.matmul(hidden_layer_2_output, hidden_layer_3['weights']), hidden_layer_3['bias']))
final_output = tf.nn.relu(tf.add(tf.matmul(hidden_layer_3_output, output_layer['weights']), output_layer['bias']))

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=final_output, labels=y))
model = tf.train.AdamOptimizer().minimize(cost)

epochs = 10

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(epochs):
        epoch_loss = 0
        for _ in range(mnist.train.num_examples/batch_size):
            P,Q = mnist.train.next_batch(batch_size)
            _,c = sess.run([model, cost], feed_dict = {x:P, y:Q})
            epoch_loss+=c

        print("Epoch no:",i,"Epoch_loss:",epoch_loss)
    correct = tf.equal(tf.argmax(final_output,1), tf.argmax(y,1))
    accuracy = tf.reduce_mean(tf.cast(correct, 'float'))    
    print("accuracy: ",accuracy.eval({x:mnist.test.images, y:mnist.test.labels}))

生成的结果是

Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz
('Epoch no:', 0, 'Epoch_loss:', 265771.25100541115)
('Epoch no:', 1, 'Epoch_loss:', 1310.440309047699)
('Epoch no:', 2, 'Epoch_loss:', 1262.8069067001343)
('Epoch no:', 3, 'Epoch_loss:', 1262.8069069385529)
('Epoch no:', 4, 'Epoch_loss:', 1262.8069067001343)
('Epoch no:', 5, 'Epoch_loss:', 1262.8069069385529)
('Epoch no:', 6, 'Epoch_loss:', 1262.8069067001343)
('Epoch no:', 7, 'Epoch_loss:', 1262.8069067001343)
('Epoch no:', 8, 'Epoch_loss:', 1262.8069064617157)
('Epoch no:', 9, 'Epoch_loss:', 1262.8069064617157)
('accuracy: ', 0.1008)

您能否告诉我此代码中结果不准确的可能原因以及如何改进？

Answer 1

您的代码存在几个问题:

1. 删除 Final_output 上的 relu 激活。 softmax_cross_entropy_with_logits 将对您的 Final_output 应用 softmax 激活。

   final_output = tf.add(tf.matmul(hidden_layer_3_output, output_layer['weights']), output_layer['bias']) 
   

2. 将权重的标准差设置为较低的值。

   'weights': tf.Variable(tf.random_normal(shape = [784, n_hidden_layer_1], stddev=0.005))