tensorflow实现AlexNet

AlexNet网络在2012年ImageNet比赛中取得最好成绩，简单阅读论文，下面是一些AlexNet网络的细节

5个卷积层，3个全连接层，训练时间为5到6天。

数据使ImageNet

15 million的标注数据，分为22w个类别

用的ImageNet的子集，训练集1.2million；验证集50w;测试集15w

ImageNet的图片的分辨率是可变的，alexnet模型把图片都转为256*256

激活函数：RELU

Pooling:OverlappingPooling

网络结构：

2/4/5卷积层只关联到同一个GPU的前一层

3卷积层关联到整个第二层（两个GPU都关联）

全连接层也都关联到前面的整个层（两个GPU都关联）

Max-pooling在1/2/5层

参数:60million

减少过拟合：

Data Augmentation

Dropout 0.5

训练的细节：

随机梯度下降

Batch:128

Momentun :0.9

weightdecay:0.0005

代码：

 

#encoding:utf-8
import sys
reload(sys)
sys.setdefaultencoding('utf-8')

import tensorflow as tf
import os
#os.environ["CNDA_VISIABLE_DEVICES"] = 3

import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot = True)
# 超参数设置
learning_rate = 0.001
training_iters = 200000
batch_size = 64
display_step = 20

#定义网络参数
n_input = 784
n_classes = 10

input_x = tf.placeholder(tf.float32, [None, n_input])
input_y = tf.placeholder(tf.float32, [None, n_classes])
keep_prob = tf.placeholder(tf.float32)

# 卷积操作
def conv2d(name, l_input, w, b):
	return tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(l_input, w, strides = [1, 1, 1, 1], padding = 'VALID'), b), name = name)

#最大下采样
def max_pool(name, l_input, k):
	return tf.nn.max_pool(l_input, ksize = [1, k, k, 1], strides = [1, k, k, 1], padding = 'VALID', name = name)

#向量转为矩阵
x = tf.reshape(input_x, shape = [-1, 28, 28, 1])

l2_loss = 0
#第一层
conv1_w = tf.Variable(tf.random_normal([3, 3, 1, 64]))
conv1_b = tf.Variable(tf.random_normal([64]))
l2_loss += tf.nn.l2_loss(conv1_w)
l2_loss += tf.nn.l2_loss(conv1_b)
conv1 = conv2d('conv1', x, conv1_w, conv1_b)
pool1 = max_pool('pool1', conv1, k = 2)
norm1 =  tf.nn.dropout(pool1, keep_prob)

#第二层
conv2_w = tf.Variable(tf.random_normal([4, 4, 64, 128]))
conv2_b = tf.Variable(tf.random_normal([128]))
l2_loss += tf.nn.l2_loss(conv2_w)
l2_loss += tf.nn.l2_loss(conv2_b)
conv2 = conv2d("conv2", norm1, conv2_w, conv2_b)
pool2 = max_pool('pool1', conv2, k = 2)
norml2 = tf.nn.dropout(pool2, keep_prob)


# 第三层
# conv3_w = tf.Variable(tf.random_normal([3, 3, 128, 256]))
# conv3_b = tf.Variable(tf.random_normal([256]))
# l2_loss += tf.nn.l2_loss(conv3_w)
# l2_loss += tf.nn.l2_loss(conv3_b)
# conv3 = conv2d('conv3', norm2, conv3_w, conv3_b)
# pool3 = max_pool('pool3', conv3, k=2)
# norm3 = tf.nn.dropout(pool3, keep_prob)
# print "norm3 shape: ", norm3.get_shape()

#全连接层，先把特征图转为向量
dense1 = tf.reshape(norml2, [-1, 5*5*128])

mlp1_w = tf.Variable(tf.random_normal([5*5*128, 1024]))
mlp1_b = tf.Variable(tf.random_normal([1024]))
l2_loss += tf.nn.l2_loss(mlp1_w)
l2_loss += tf.nn.l2_loss(mlp1_b)
dense1 = tf.nn.relu(tf.matmul(dense1, mlp1_w) + mlp1_b, name = 'fc1')

mlp2_w = tf.Variable(tf.random_normal([1024, 1024]))
mlp2_b = tf.Variable(tf.random_normal([1024]))
l2_loss += tf.nn.l2_loss(mlp2_w)
l2_loss += tf.nn.l2_loss(mlp2_b)
dense2 = tf.nn.relu(tf.matmul(dense1, mlp2_w) + mlp2_b, name = 'fc2')

out_w = tf.Variable(tf.random_normal([1024, n_classes]))
out_b = tf.Variable(tf.random_normal([n_classes]))
l2_loss += tf.nn.l2_loss(out_w)
l2_loss += tf.nn.l2_loss(out_b)
pred_out = tf.matmul(dense2, out_w) + out_b

#定义损失函数和学习步骤
cost  = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = pred_out, labels = input_y))
cost += 0.1 * l2_loss
optimzier = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost)

#测试网络
correct_pred = tf.equal(tf.argmax(pred_out, 1), tf.argmax(input_y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

#模型执行过程-训练过程
with tf.Session() as sess:
	init = tf.global_variables_initializer()
	sess.run(init)
	step = 1
	# Keep Training until reach max iterations
	while step * batch_size < training_iters:
		batch_xs, batch_ys = mnist.train.next_batch(batch_size)
		sess.run(optimzier, feed_dict = {input_x : batch_xs, input_y : batch_ys, keep_prob : 0.5})
		if step % display_step == 0:
			#计算精度
			acc, loss = sess.run([accuracy, cost], {input_x : batch_xs, input_y : batch_ys, keep_prob : 1})
			print "Iter:" + str(step*batch_size) + ",Minibatch Loss=" + "{:.6f}".format(loss)+ ",Training Accuracy = " + "{:.5f}".format(acc)
		step += 1
	print "Optimization Finished!"
	print "Test Accuracy:", sess.run(accuracy, feed_dict = {input_x : mnist.test_images[:256], input_y : mnist.test_labels[:256], keep_prob : 1})