LeNet训练MNIST

jupyter notebook： https://github.com/Penn000/NN/blob/master/jupyter/LeNet/LeNet.ipynb

LeNet训练MNIST

1 import warnings
2 warnings.filterwarnings('ignore')  # 不打印 warning 
3 
4 import tensorflow as tf
5 import numpy as np
6 import os

加载MNIST数据集

分别加载MNIST训练集、测试集、验证集

1 from tensorflow.examples.tutorials.mnist import input_data
2 
3 mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
4 X_train, y_train = mnist.train.images, mnist.train.labels
5 X_test, y_test = mnist.test.images, mnist.test.labels
6 X_validation, y_validation = mnist.validation.images, mnist.validation.labels

1 print("Image Shape: {}".format(X_train.shape))
2 print("label Shape: {}".format(y_train.shape))
3 print()
4 print("Training Set:   {} samples".format(len(X_train)))
5 print("Validation Set: {} samples".format(len(X_validation)))
6 print("Test Set:       {} samples".format(len(X_test)))

数据处理

由于LeNet的输入为32x32xC（C为图像通道数），而MNIST每张图像的尺寸为28x28，所以需要对图像四周进行填充，并添加一维，使得每幅图像的形状为32x32x1。

1 # 使用0对图像四周进行填充
2 X_train = np.array([np.pad(X_train[i].reshape((28, 28)), (2, 2), 'constant')[:, :, np.newaxis] for i in range(len(X_train))])
3 X_validation = np.array([np.pad(X_validation[i].reshape((28, 28)), (2, 2), 'constant')[:, :, np.newaxis] for i in range(len(X_validation))])
4 X_test = np.array([np.pad(X_test[i].reshape((28, 28)), (2, 2), 'constant')[:, :, np.newaxis] for i in range(len(X_test))])
5     
6 print("Updated Image Shape: {}".format(X_train.shape))

MNIST数据展示

 1 import random
 2 import numpy as np
 3 import matplotlib.pyplot as plt
 4 %matplotlib inline
 5 
 6 index = random.randint(0, len(X_train))
 7 image = X_train[index].squeeze().reshape((32, 32))
 8 
 9 plt.figure(figsize=(2,2))
10 plt.imshow(image, cmap="gray")
11 print(y_train[index])

LeNet网络结构

 

Input

The LeNet architecture accepts a 32x32xC image as input, where C is the number of color channels. Since MNIST images are grayscale, C is 1 in this case. LeNet的输入为32x32xC的图像，C为图像的通道数。在MNIST中，图像为灰度图，因此C等于1。

Architecture

Layer 1: Convolutional. 输出为28x28x6的张量。

Activation. 激活函数。

Pooling. 输出为14x14x6的张量。

Layer 2: Convolutional. 输出为10x10x16的张量。

Activation. 激活函数。

Pooling. 输出为5x5x16的张量。

Flatten. 将张量展平为一维向量，使用tf.contrib.layers.flatten可以实现。

Layer 3: Fully Connected. 输出为120长度的向量。

Activation. 激活函数。

Layer 4: Fully Connected. 输出为84长度的向量。

Activation. 激活函数。

Layer 5: Fully Connected (Logits). 输出为10长度的向量。

1 # 卷积层
2 def conv_layer(x, filter_shape, stride, name):
3     with tf.variable_scope(name):
4         W = tf.get_variable('weights', shape=filter_shape, initializer=tf.truncated_normal_initializer())
5         b = tf.get_variable('biases', shape=filter_shape[-1], initializer=tf.zeros_initializer())
6     return tf.nn.conv2d(x, W, strides=stride, padding='VALID', name=name) + b

1 # 全连接层
2 def fc_layer(x, in_size, out_size, name):
3     with tf.variable_scope(name):
4         W = tf.get_variable('weights', shape=(in_size, out_size), initializer=tf.truncated_normal_initializer())
5         b = tf.get_variable('biases', shape=(out_size), initializer=tf.zeros_initializer())
6     
7     return tf.nn.xw_plus_b(x, W, b, name=name)

1 def relu_layer(x, name):
2     return tf.nn.relu(x, name=name)

 1 from tensorflow.contrib.layers import flatten
 2 
 3 def LeNet(x): 
 4     conv1 = conv_layer(x, filter_shape=(5, 5, 1, 6), stride=[1, 1, 1, 1], name='conv1')
 5     relu1 = relu_layer(conv1, 'relu1')
 6     max_pool1 = max_pool_layer(relu1,  kernel_size=[1, 2, 2, 1], stride=[1, 2, 2, 1], name='max_pool1')
 7     
 8     conv2 = conv_layer(max_pool1, filter_shape=(5, 5, 6, 16), stride=[1, 1, 1, 1], name='conv2')
 9     relu2 = relu_layer(conv2, 'relu2')
10     max_pool2 = max_pool_layer(relu2,  kernel_size=[1, 2, 2, 1], stride=[1, 2, 2, 1], name='max_pool1')
11     
12     flat = flatten(max_pool2)
13     
14     fc3 = fc_layer(flat, 400, 120, name='fc3')
15     relu3 = relu_layer(fc3, 'relu3')
16     
17     fc4 = fc_layer(relu3, 120, 84, name='fc4')
18     relu4 = relu_layer(fc4, 'relu4')
19     
20     logits = fc_layer(relu4, 84, 10, name='fc5')
21     
22     return logits

TensorFlow设置

 1 EPOCHS = 10
 2 BATCH_SIZE = 128
 3 log_dir = './log/'
 4 
 5 x = tf.placeholder(tf.float32, (None, 32, 32, 1))
 6 y = tf.placeholder(tf.int32, (None, 10))
 7 
 8 # 定义损失函数
 9 logits = LeNet(x)
10 cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=logits)
11 loss = tf.reduce_mean(cross_entropy)
12 train = tf.train.AdamOptimizer(learning_rate=0.01).minimize(loss)

训练

 1 from sklearn.utils import shuffle
 2 import shutil
 3 log_dir = './logs/'
 4 if os.path.exists(log_dir):
 5     shutil.rmtree(log_dir)
 6 os.makedirs(log_dir)
 7 train_writer = tf.summary.FileWriter(log_dir+'train/')
 8 valid_writer = tf.summary.FileWriter(log_dir+'valid/')
 9 
10 ckpt_path = './ckpt/'
11 saver = tf.train.Saver()
12 
13 with tf.Session() as sess:
14     sess.run(tf.global_variables_initializer())
15     n_samples = len(X_train)
16     
17     step = 0
18     for i in range(EPOCHS):
19         X_train, y_train = shuffle(X_train, y_train) # 打乱数据
20         # 使用mini-batch训练
21         for offset in range(0, n_samples, BATCH_SIZE):
22             end = offset + BATCH_SIZE
23             batch_x, batch_y = X_train[offset:end], y_train[offset:end]
24             sess.run(train, feed_dict={x: batch_x, y: batch_y})
25             
26             train_loss = sess.run(loss, feed_dict={x: batch_x, y: batch_y})
27             train_summary = tf.Summary(value=[
28                 tf.Summary.Value(tag="loss", simple_value=train_loss)
29             ])
30             train_writer.add_summary(train_summary, step)
31             train_writer.flush()
32             step += 1
33         
34         # 每个epoch使用验证集对网络进行验证
35         valid_loss = sess.run(loss, feed_dict={x: X_validation, y: y_validation})
36         valid_summary = tf.Summary(value=[
37                 tf.Summary.Value(tag="loss", simple_value=valid_loss)
38         ])
39         valid_writer.add_summary(valid_summary, step)
40         valid_writer.flush()
41         
42         print('epoch', i, '>>> loss:', valid_loss)
43     
44     # 保存模型
45     saver.save(sess, ckpt_path + 'model.ckpt')
46     print("Model saved")

训练和验证的loss曲线

测试

1 correct = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))
2 accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
3 
4 with tf.Session() as sess:
5     saver.restore(sess, tf.train.latest_checkpoint('./ckpt'))
6 
7     test_accuracy = sess.run(accuracy, feed_dict={x: X_test, y: y_test})
8     print("Test Accuracy = {}".format(test_accuracy))