VGGNet 实现

import tensorflow as tf
from datetime import datetime
import math
import time

batch_size = 12
num_batches = 100

def myprint(conv):
	# print(conv.op.name, ' ', conv.get_shape().as_list())
	str1 = str(conv.op.name) + ' ' + str(conv.get_shape().as_list()) + "\n"
	f = open('test.txt', 'a')  # 若是'wb'就表示写二进制文件
	f.write(str1)
	f.close()

def conv_op(input, name, kernel_h, kernel_w, num_out, step_h, step_w, para):
	# 获取输入图像的通道
	init = tf.contrib.layers.xavier_initializer_conv2d()
	num_in = input.get_shape()[-1].value

	with tf.name_scope(name) as scope:
		kernel = tf.get_variable(scope + "w" , shape=[kernel_h, kernel_w, num_in, num_out], dtype=tf.float32, initializer=init)
		conv = tf.nn.conv2d(input, kernel, (1, step_h, step_w,1),padding="SAME")
		biases = tf.Variable(tf.constant(0.0, shape=[num_out], dtype=tf.float32), trainable=True, name="b")
		avtivation = tf.nn.relu(tf.nn.bias_add(conv, biases), name = scope)

		para += [kernel, biases]
		return avtivation

def fc_op(input, name, num_out, para):
	init = tf.contrib.layers.xavier_initializer_conv2d()
	num_in = input.get_shape()[-1].value
	with tf.name_scope(name) as scope:
		weights = tf.get_variable(scope + "w", shape=[num_in, num_out], dtype= tf.float32, initializer=init)
		biases = tf.Variable(tf.constant(0.1, shape=[num_out], dtype=tf.float32), name="b")

		avtivation = tf.nn.relu_layer(input, weights, biases)

		para += [weights, biases]
		return avtivation

# 图像的数据输入是224 * 224 * 3
def inference_op(input, keep_prob):
	parameters = []
# 	第一段卷积, 输出的大小是112 * 112 * 64
	conv1_1 = conv_op(input, name="conv1_1", kernel_h=3, kernel_w=3, num_out=4, step_h=1, step_w=1, para=parameters)
	myprint(conv1_1)
	conv1_2 = conv_op(conv1_1, name="conv1_2", kernel_h=3, kernel_w=3, num_out=64, step_h=1, step_w=1, para=parameters)
	myprint(conv1_2)
	pool1 = tf.nn.max_pool(conv1_2, ksize=[1,2,2,1], strides=[1,2,2,1], padding="SAME", name="pool1")
	myprint(pool1)

# 	第二段卷积,输出的大小是56*56*128
	conv2_1 = conv_op(pool1, name = "conv2_1", kernel_h=3, kernel_w=3, num_out=128,step_h=1, step_w=1, para=parameters)
	myprint(conv2_1)
	conv2_2 = conv_op(conv2_1, name = "conv2_2", kernel_h=3, kernel_w=3, num_out=128,step_h=1, step_w=1, para=parameters)
	myprint(conv2_2)
	pool2 = tf.nn.max_pool(conv2_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME", name="pool2")
	myprint(pool2)

# 第三段卷积, 输出的大小是28*28*256
	conv3_1 = conv_op(pool2, name="conv3_1", kernel_h=3, kernel_w=3, num_out=256, step_h=1, step_w=1, para=parameters)
	myprint(conv3_1)
	conv3_2 = conv_op(conv3_1, name="conv3_2", kernel_h=3, kernel_w=3, num_out=256, step_h=1, step_w=1, para=parameters)
	myprint(conv3_2)
	conv3_3 = conv_op(conv3_2, name="conv3_3", kernel_h=3, kernel_w=3, num_out=256, step_h=1, step_w=1, para=parameters)
	myprint(conv3_3)
	pool3 = tf.nn.max_pool(conv3_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME", name="pool3")
	myprint(pool3)

# 	第四段卷积, 输出的大小是14*14*512
	conv4_1 = conv_op(pool3, name="conv4_1", kernel_h=3, kernel_w=3, num_out=512, step_h=1, step_w=1, para=parameters)
	myprint(conv4_1)
	conv4_2 = conv_op(conv4_1, name="conv4_2", kernel_h=3, kernel_w=3, num_out=512, step_h=1, step_w=1, para=parameters)
	myprint(conv4_2)
	conv4_3 = conv_op(conv4_2, name="conv4_3", kernel_h=3, kernel_w=3, num_out=512, step_h=1, step_w=1, para=parameters)
	myprint(conv4_3)
	pool4 = tf.nn.max_pool(conv4_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME", name="pool4")
	myprint(pool4)

#   第五段卷积, 输出的大小是7*7*512
	conv5_1 = conv_op(pool4, name="conv5_1", kernel_h=3, kernel_w=3, num_out=512, step_h=1, step_w=1, para=parameters)
	myprint(conv5_1)
	conv5_2 = conv_op(conv5_1, name="conv5_2", kernel_h=3, kernel_w=3, num_out=512, step_h=1, step_w=1, para=parameters)
	myprint(conv5_2)
	conv5_3 = conv_op(conv5_2, name="conv5_3", kernel_h=3, kernel_w=3, num_out=512, step_h=1, step_w=1, para=parameters)
	myprint(conv5_3)
	pool5 = tf.nn.max_pool(conv5_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME", name="pool5")
	myprint(pool5)

	pool_shape = pool5.get_shape().as_list()
	flattened_shape = pool_shape[1]*pool_shape[2]*pool_shape[3]
	reshaped = tf.reshape(pool5, [-1, flattened_shape], name="reshped")
	myprint(reshaped)

# 	创建第一个全连层
	fc_6 = fc_op(reshaped, name="fc6", num_out=4096, para=parameters)
	fc_6_drop = tf.nn.dropout(fc_6, keep_prob, name="fc_6_drop")
	myprint(fc_6_drop)

# 	创建第二个全连层
	fc_7 = fc_op(fc_6_drop, name="fc7", num_out=4096, para=parameters)
	fc_7_drop = tf.nn.dropout(fc_7, keep_prob, name="fc_7_drop")
	myprint(fc_7_drop)
# 	创建第三个全连层
	fc_8 = fc_op(fc_7_drop, name="fc8", num_out=1000, para=parameters)
	softmax = tf.nn.softmax(fc_8)
	myprint(fc_8)
	myprint(softmax)

# 	得到结果
	predictions = tf.argmax(softmax, 1)
	return predictions, softmax, fc_8, parameters



with tf.Graph().as_default():
	image_size = 224
	images = tf.Variable(tf.random_normal([batch_size, image_size, image_size, 3], dtype=tf.float32, stddev=1e-1))

	keep_prob = tf.placeholder(tf.float32)

	predictions, softmax, fc_8, parameters = inference_op(images, keep_prob)

	init_op = tf.global_variables_initializer()

	# 使用BFC算法确定GPU内存最佳分配策略
	config = tf.ConfigProto()
	config.gpu_options.allocator_type = "BFC"

	with tf.Session(config=config) as sess:
		sess.run(init_op)
		num_steps_burn_in = 10

		total_dura = 0.0
		total_dura_squared = 0.0

		back_total_dura = 0.0
		back_total_dura_squared = 0.0

		for i in range(num_batches + num_steps_burn_in):
			start_time = time.time()
			_ = sess.run(predictions, feed_dict={keep_prob: 1.0})
			duration = time.time() - start_time
			if i >= num_steps_burn_in:
				if i % 10 == 0:
					print("%s: step %d, duration = %.3f" %
					      (datetime.now(), i - num_steps_burn_in, duration))
				total_dura += duration
				total_dura_squared += duration * duration
		average_time = total_dura / num_batches

		print("%s: Forward across %d steps, %.3f +/- %.3f sec / batch" %
		      (datetime.now(), num_batches, average_time,
		       math.sqrt(total_dura_squared / num_batches - average_time * average_time)))

		# 定义求解梯度的操作
		grad = tf.gradients(tf.nn.l2_loss(fc_8), parameters)

		# 运行反向传播测试过程
		for i in range(num_batches + num_steps_burn_in):
			start_time = time.time()
			_ = sess.run(grad, feed_dict={keep_prob: 0.5})
			duration = time.time() - start_time
			if i >= num_steps_burn_in:
				if i % 10 == 0:
					print("%s: step %d, duration = %.3f" %
					      (datetime.now(), i - num_steps_burn_in, duration))
				back_total_dura += duration
				back_total_dura_squared += duration * duration
		back_avg_t = back_total_dura / num_batches

		# 打印反向传播的运算时间信息
		print("%s: Forward-backward across %d steps, %.3f +/- %.3f sec / batch" %
		      (datetime.now(), num_batches, back_avg_t,
		       math.sqrt(back_total_dura_squared / num_batches - back_avg_t * back_avg_t)))