使用tensorflow构建如上图所示的CNN用于对MNIST数据集进行softmax classification。
理论部分不再赘述，完整的代码如下：
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
# hyperparameter
learning_rate = 0.001
training_epoches = 20
batch_size = 100
class Model:
def __init__(self, sess, name):
self.sess = sess
self.name = name
self._build_net()
def _build_net(self):
# with tf.variable_scope(self.name):
self.training = tf.placeholder(tf.bool)
# input placeholder for X & Y
self.X = tf.placeholder(tf.float32, [None, 784])
self.Y = tf.placeholder(tf.float32, [None, 10])
# img 28x28x1 (black/white)
X_img = tf.reshape(self.X, [-1, 28, 28, 1])
# convolutional layer 1 & pooling layer 1
conv1 = tf.layers.conv2d(inputs=X_img, filters=32, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2],
padding="SAME", strides=2)
dropout1 = tf.layers.dropout(inputs=pool1, rate=0.3, training=self.training)
# convolutional layer 2 & pooling layer 2
conv2 = tf.layers.conv2d(inputs=dropout1, filters=64, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2],
padding="SAME", strides=2)
dropout2 = tf.layers.dropout(inputs=pool2, rate=0.3, training=self.training)
# convolutional layer 3 & pooling layer 3
conv3 = tf.layers.conv2d(inputs=dropout2, filters=128, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
pool3 = tf.layers.max_pooling2d(inputs=conv3, pool_size=[2, 2],
padding="SAME", strides=2)
dropout3 = tf.layers.dropout(inputs=pool3, rate=0.3, training=self.training)
# dense layer with Relu
flat = tf.reshape(dropout3, [-1, 128 * 4 * 4])
dense4 = tf.layers.dense(inputs=flat, units=625, activation=tf.nn.relu)
dropout4 = tf.layers.dropout(inputs=dense4, rate=0.5, training=self.training)
# FC layer 625 input -> 10 output, no activation function
self.logits = tf.layers.dense(inputs=dropout4, units=10)
# define loss & optimizer
self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(
logits = self.logits, labels=self.Y))
self.optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(self.cost)
# accuracy
correct_prediction = tf.equal(tf.argmax(self.logits, 1), tf.argmax(self.Y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
def train(self, x_data, y_data, training=True):
return self.sess.run([self.cost, self.optimizer],
feed_dict={self.X: x_data, self.Y: y_data, self.training: training})
def predict(self, x_test, training=False):
return self.sess.run(self.logits,
feed_dict={self.X :x_test, self.training: training})
def get_accuracy(self, x_test, y_test, training=False):
return self.sess.run(self.accuracy,
feed_dict={self.X: x_test,self.Y: y_test, self.training: training})
# train the models
with tf.Session() as sess:
models = []
num_models = 2
for m in range(num_models):
models.append(Model(sess, "modal"+str(m)))
sess.run(tf.global_variables_initializer())
print('Learning Start!')
for epoch in range(training_epoches):
avg_cost_list = np.zeros(len(models))
total_batch = int(mnist.train.num_examples / batch_size)
for i in range(total_batch):
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
# train each modal
for m_id, m in enumerate(models):
c, _ = m.train(batch_xs, batch_ys)
avg_cost_list[m_id] += c / total_batch
print('Epoch: ', '%04d' %(epoch+1), 'cost=', avg_cost_list)
print('Learning finished!')
# test & accuracy
test_size = len(mnist.test.labels)
predictions = np.zeros([test_size, 10])
for m_id,m in enumerate(models):
print(m_id, "Accuracy:", m.get_accuracy(mnist.test.images, mnist.test.labels))
p = m.predict(mnist.test.images)
predictions += p
ensemble_correct_prediction = tf.equal(tf.argmax(predictions, 1), tf.argmax(mnist.test.labels, 1))
ensemble_accuracy = tf.reduce_mean(tf.cast(ensemble_correct_prediction, tf.float32))
print("Ensemble_accuracy:", sess.run(ensemble_accuracy))
结果：
Learning Start!
Epoch: 0001 cost= [0.29211415 0.28355632]
Epoch: 0002 cost= [0.08716567 0.0870499 ]
Epoch: 0003 cost= [0.06902521 0.06623169]
Epoch: 0004 cost= [0.05563359 0.05452387]
Epoch: 0005 cost= [0.04963774 0.04871382]
Epoch: 0006 cost= [0.04462749 0.04449957]
Epoch: 0007 cost= [0.04132144 0.03907955]
Epoch: 0008 cost= [0.03792324 0.03861412]
Epoch: 0009 cost= [0.0354344 0.03323769]
Epoch: 0010 cost= [0.03516847 0.03405525]
Epoch: 0011 cost= [0.03143759 0.03219781]
Epoch: 0012 cost= [0.03051504 0.02993162]
Epoch: 0013 cost= [0.02906878 0.02711077]
Epoch: 0014 cost= [0.02729127 0.02754832]
Epoch: 0015 cost= [0.02729633 0.02632647]
Epoch: 0016 cost= [0.02438517 0.02701174]
Epoch: 0017 cost= [0.02482958 0.0244114 ]
Epoch: 0018 cost= [0.02455271 0.02649499]
Epoch: 0019 cost= [0.02371975 0.02178147]
Epoch: 0020 cost= [0.02260135 0.0213784 ]
Learning finished!
0 Accuracy: 0.995
1 Accuracy: 0.9949
Ensemble_accuracy: 0.9954
结果前面的其实有很长的warning，这里没有给出。warning是说新版本的tensorflow把mnist数据集移动到了别的地方，建议你从别的地方导入进来。这篇博文仅做例子。实际使用tensorflow的时候，你都是自己写读取数据的函数什么的，需要根据数据集的存储格式写不同的Python代码。