這是LeNet-5模型一個簡版
第一層輸入層是28x28x1的圖片
第二層卷積層採用32個5x5的fiter，邊緣用全0填補，得到28x28x32的輸出
第三層是池化層 採用2x2的fiter，步長也是【2,2】，得到輸出14x14x28的圖片
第四層卷積採用5x5x64的fiter，邊緣用全零填補，得到14x14x64的輸出
第五層池化跟第二層一樣，得到7x7x64的輸出
第六層全連接層，將7x7x64=3236個節點作爲輸入，輸出層節點512個
第七層，全連接層 ，輸出層是10個節點
以下是代碼和運行結果
LeNet-5.py
import tensorflow as tf
input_node = 784
output_node = 10
image_size = 28
num_labels = 10
conv1_deep = 32
conv1_size = 5
conv2_deep = 64
conv2_size = 5
fc_size = 512
def inference(input_tensor, train, regularizer):
init2 = tf.constant_initializer(0.0)
with tf.variable_scope('layer1-conv1'):
conv1_weights = tf.get_variable('weight', [conv1_size, conv1_size, 1, conv1_deep], initializer=tf.truncated_normal_initializer(stddev=0.1)) # 5*5*1*32
conv1_biases = tf.get_variable('bias', [conv1_deep], initializer=init2)
conv1 = tf.nn.conv2d(input_tensor, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
with tf.variable_scope('layer2-pool1'):
pool1 = tf.nn.max_pool(relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
with tf.variable_scope('layer3-conv2'):
conv2_weights = tf.get_variable('weight', [conv2_size, conv2_size, conv1_deep, conv2_deep],initializer=tf.truncated_normal_initializer(stddev=0.1))
conv2_biases = tf.get_variable('bias', [conv2_deep], initializer=init2)
conv2 = tf.nn.conv2d(pool1, conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
with tf.variable_scope('layer4-pool2'):
pool2 = tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool_shape = pool2.get_shape().as_list()
nodes = pool_shape[1] * pool_shape[2] * pool_shape[3]
reshape = tf.reshape(pool2, [pool_shape[0], nodes])
with tf.variable_scope('layer5-fc1'):
fc1_weights = tf.get_variable('weight', [nodes, fc_size], initializer=tf.truncated_normal_initializer(stddev=0.1))
if regularizer != None:
tf.add_to_collection('losses', regularizer(fc1_weights))
fc1_biases = tf.get_variable('bias', [fc_size], initializer=init2)
fc1 = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
if train: fc1 = tf.nn.dropout(fc1, 0.5)
with tf.variable_scope('layer6-fc2'):
fc2_weights = tf.get_variable('weight', [fc_size, num_labels],initializer=tf.truncated_normal_initializer(stddev=0.1))
if regularizer != None:
tf.add_to_collection('losses', regularizer(fc2_weights))
fc2_biases = tf.get_variable('bias', [num_labels], initializer=init2)
logit = tf.matmul(fc1, fc2_weights) + fc2_biases
return logit
MNIST_train.py
import os
import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
# 加載LeNet_5.py中定義的常量和前向傳播的函數
import LeNet_5
# 配置神經網絡的參數
BATCH_SIZE = 100
LEARNING_RATE_BASE = 0.01
LEARNING_RATE_DECAY = 0.99
REGULARAZTION_RATE = 0.0001
TRAINING_STEPS = 30000
MOVING_AVERAGE_DECAY = 0.99
# 模型保存的路徑和文件名
MODEL_SAVE_PATH = "model/"
MODEL_NAME = "model.ckpt"
def train(mnist):
# 定義輸入輸出placeholder
# 調整輸入數據placeholder的格式，輸入爲一個四維矩陣
x = tf.placeholder(tf.float32, [
BATCH_SIZE, # 第一維表示一個batch中樣例的個數
LeNet_5.image_size, # 第二維和第三維表示圖片的尺寸
LeNet_5.image_size,
1], # 第四維表示圖片的深度，對於RBG格式的圖片，深度爲3
name='x-input')
y_ = tf.placeholder(tf.float32, [None, LeNet_5.output_node], name='y-input')
regularizer = tf.contrib.layers.l2_regularizer(REGULARAZTION_RATE)
# 直接使用LeNet_5.py中定義的前向傳播過程
y = LeNet_5.inference(x, True, regularizer)
global_step = tf.Variable(0, trainable=False)
#定義損失函數、學習率、滑動平均操作以及訓練過程
variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
variable_averages_op = variable_averages.apply(tf.trainable_variables())
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
cross_entropy_mean = tf.reduce_mean(cross_entropy)
loss = cross_entropy_mean + tf.add_n(tf.get_collection('losses'))
learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, global_step, mnist.train.num_examples/BATCH_SIZE, LEARNING_RATE_DECAY)
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
sum=tf.reduce_sum(tf.cast(tf.equal(tf.argmax(y,1),tf.argmax(y_,1)),dtype=tf.int32))
with tf.control_dependencies([train_step, variable_averages_op]):
train_op = tf.no_op(name='train')
saver = tf.train.Saver()
with tf.Session() as sess:
tf.global_variables_initializer().run()
# 驗證和測試的過程將會有一個獨立的程序來完成
for i in range(TRAINING_STEPS):
xs, ys = mnist.train.next_batch(BATCH_SIZE)
#類似地將輸入的訓練數據格式調整爲一個四維矩陣，並將這個調整後的數據傳入sess.run過程
reshaped_xs = np.reshape(xs, (BATCH_SIZE, LeNet_5.image_size, LeNet_5.image_size, 1))
_, loss_value, step = sess.run([train_op, loss, global_step], feed_dict={x: reshaped_xs, y_: ys})
#每1000輪保存一次模型。
if i%200 == 0:
# 輸出當前的訓練情況。這裏只輸出了模型在當前訓練batch上的損失函數大小。通過損失函數的大小可以大概瞭解訓練的情況。
# 在驗證數據集上的正確率信息會有一個單獨的程序來生成。
print("After %d training step(s), loss on training batch is %f." % (step, loss_value))
SUM=0;
for j in range(mnist.validation.num_examples//BATCH_SIZE):
xs_,ys_=mnist.validation.next_batch(BATCH_SIZE)
rx=np.reshape(xs_,(BATCH_SIZE,LeNet_5.image_size,LeNet_5.image_size,1))
SUM+=sess.run([sum],feed_dict={x:rx,y_:ys_})[0]
print("After %d training step(s), corrcet rate is %f." % (step, SUM/mnist.validation.num_examples))
saver.save(sess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step=global_step)
def main(argv=None):
mnist = input_data.read_data_sets("C:/Users/tang/Desktop/deeplearning/mnist數據集", one_hot=True)
train(mnist)
if __name__ == '__main__':
tf.app.run()
大概運行了1個小時，其運行結果如下：
After 1 training step(s), loss on training batch is 6.577364.
After 1 training step(s), corrcet rate is 0.112800.
After 201 training step(s), loss on training batch is 1.134218.
After 201 training step(s), corrcet rate is 0.820400.
After 401 training step(s), loss on training batch is 1.095968.
After 401 training step(s), corrcet rate is 0.882800.
After 601 training step(s), loss on training batch is 0.928677.
After 601 training step(s), corrcet rate is 0.908400.
After 801 training step(s), loss on training batch is 0.948598.
After 801 training step(s), corrcet rate is 0.921400.
After 1001 training step(s), loss on training batch is 0.963806.
After 1001 training step(s), corrcet rate is 0.929200.
After 1201 training step(s), loss on training batch is 0.758020.
After 1201 training step(s), corrcet rate is 0.937800.
After 1401 training step(s), loss on training batch is 0.855776.
After 1401 training step(s), corrcet rate is 0.944600.
After 1601 training step(s), loss on training batch is 0.775362.
After 1601 training step(s), corrcet rate is 0.948600.
After 1801 training step(s), loss on training batch is 0.752938.
After 1801 training step(s), corrcet rate is 0.954200.
After 2001 training step(s), loss on training batch is 0.873272.
After 2001 training step(s), corrcet rate is 0.953400.
After 2201 training step(s), loss on training batch is 0.688549.
After 2201 training step(s), corrcet rate is 0.956200.
After 2401 training step(s), loss on training batch is 0.810790.
After 2401 training step(s), corrcet rate is 0.959600.
After 2601 training step(s), loss on training batch is 0.817977.
After 2601 training step(s), corrcet rate is 0.958400.
After 2801 training step(s), loss on training batch is 0.754317.
After 2801 training step(s), corrcet rate is 0.960000.
After 3001 training step(s), loss on training batch is 0.734173.
After 3001 training step(s), corrcet rate is 0.962600.
After 3201 training step(s), loss on training batch is 0.665527.
After 3201 training step(s), corrcet rate is 0.963800.
After 3401 training step(s), loss on training batch is 0.674073.
After 3401 training step(s), corrcet rate is 0.966200.
After 3601 training step(s), loss on training batch is 0.828704.
After 3601 training step(s), corrcet rate is 0.968400.
After 3801 training step(s), loss on training batch is 0.803491.
After 3801 training step(s), corrcet rate is 0.966400.
After 4001 training step(s), loss on training batch is 0.660605.
After 4001 training step(s), corrcet rate is 0.965200.
After 4201 training step(s), loss on training batch is 0.738579.
After 4201 training step(s), corrcet rate is 0.970400.
After 4401 training step(s), loss on training batch is 0.785027.
After 4401 training step(s), corrcet rate is 0.966800.
After 4601 training step(s), loss on training batch is 0.692350.
After 4601 training step(s), corrcet rate is 0.970400.
After 4801 training step(s), loss on training batch is 0.681506.
After 4801 training step(s), corrcet rate is 0.968600.
After 5001 training step(s), loss on training batch is 0.747684.
After 5001 training step(s), corrcet rate is 0.968400.
After 5201 training step(s), loss on training batch is 0.662478.
After 5201 training step(s), corrcet rate is 0.973200.
After 5401 training step(s), loss on training batch is 0.678184.
After 5401 training step(s), corrcet rate is 0.973200.
After 5601 training step(s), loss on training batch is 0.699767.
After 5601 training step(s), corrcet rate is 0.970800.
After 5801 training step(s), loss on training batch is 0.633660.
After 5801 training step(s), corrcet rate is 0.970400.
After 6001 training step(s), loss on training batch is 0.646770.
After 6001 training step(s), corrcet rate is 0.970000.
After 6201 training step(s), loss on training batch is 0.703785.
After 6201 training step(s), corrcet rate is 0.972200.
After 6401 training step(s), loss on training batch is 0.705623.
After 6401 training step(s), corrcet rate is 0.974800.
After 6601 training step(s), loss on training batch is 0.656697.
After 6601 training step(s), corrcet rate is 0.974600.
After 6801 training step(s), loss on training batch is 0.646331.
After 6801 training step(s), corrcet rate is 0.974200.
After 7001 training step(s), loss on training batch is 0.781423.
After 7001 training step(s), corrcet rate is 0.977000.
After 7201 training step(s), loss on training batch is 0.675177.
After 7201 training step(s), corrcet rate is 0.977800.
After 7401 training step(s), loss on training batch is 0.724536.
After 7401 training step(s), corrcet rate is 0.975400.
After 7601 training step(s), loss on training batch is 0.728857.
After 7601 training step(s), corrcet rate is 0.975800.
After 7801 training step(s), loss on training batch is 0.649602.
After 7801 training step(s), corrcet rate is 0.976200.
After 8001 training step(s), loss on training batch is 0.665193.
After 8001 training step(s), corrcet rate is 0.978400.
After 8201 training step(s), loss on training batch is 0.686455.
After 8201 training step(s), corrcet rate is 0.978000.
After 8401 training step(s), loss on training batch is 0.682056.
After 8401 training step(s), corrcet rate is 0.978600.
After 8601 training step(s), loss on training batch is 0.653046.
After 8601 training step(s), corrcet rate is 0.977400.
After 8801 training step(s), loss on training batch is 0.680877.
After 8801 training step(s), corrcet rate is 0.978000.
After 9001 training step(s), loss on training batch is 0.734092.
After 9001 training step(s), corrcet rate is 0.977400.
After 9201 training step(s), loss on training batch is 0.683104.
After 9201 training step(s), corrcet rate is 0.979000.
After 9401 training step(s), loss on training batch is 0.625013.
After 9401 training step(s), corrcet rate is 0.978200.
After 9601 training step(s), loss on training batch is 0.672356.
After 9601 training step(s), corrcet rate is 0.980400.
After 9801 training step(s), loss on training batch is 0.647361.
After 9801 training step(s), corrcet rate is 0.978000.
效果沒有達到書上說的99.4%，有待改進