TensorFlow学习－－VGGNet实现&图像识别

VGGNet架构

VGGNet结构

CNN的发展(刘昕博士的《CNN的近期进展与实用技巧》)如图：

VGGNet是由牛津大学计算机视觉组和Google DeepMind公司研发的深度卷积神经网络，通过加深网络结构提升性能．具体来说，VGGNet通过反复堆叠3*3的小卷积核和2*2的最大池化层构造了16-19层深的卷积神经网络．凭借该结构，VGGNet在ILSVRC-2014比赛上取得了分类项目的第2名和定位项目的第1名的成绩．

如图，VGGNet各级别网络结构图：

其中，网络D即VGGNet-16，网络E即VGGNet-19.
从A到E每一级网络逐渐变深．由于参数量主要消耗在最后3个全连接层上，所以网络的参数量并未随深度的加深而大量增加．
VGGNet各级别网络参数量，如图：

VGG16的宏观架构

VGG16的宏观架构：

VGGNet有5段卷积，每一段有2-3个卷积层，每段结尾由一个最大池化层来缩小图片尺寸．
每段内的卷积核数量一样，越靠后的段卷积核数量越多： 64→128→256→512→512 .其中2个3*3的卷积层串联相当于1个5*5的卷积层，即感受野大小为5*5，如图：

同样，3个3*3的卷积层串联相当于1个7*7的卷积层，但串联后的卷积层会拥有更少的参数量和更多的非线性变换，使得CNN对特征的学习能力更强．

VGG结论

VGG作者对比各级网络时总结出一下观点：

1. LRN层作用不大．
2. 越深的网络效果越好．
3. 1*1的卷积也是很有效的，但没有3*3的卷积好，大一些的卷积核可以学到更大的空间特征．

TensorFlow实现VGGNet-16

#!/usr/bin/python
# -*- coding: UTF-8 -*-

# TensorFlow实现VGGNet-16

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

# 卷积层
# kh, kw：卷积核尺寸
# n_out：卷积核输出通道数
# dh, dw：步长
# p：参数列表
def convLayer(x, name, kh, kw, n_out, dh, dw, p):
    # 输入数据的通道数
    n_in = x.get_shape()[-1].value
    # 设置scope
    with tf.name_scope(name) as scope:
        # 卷积核参数
        kernel = tf.get_variable(scope+'w',
        shape=[kh, kw, n_in, n_out], dtype=tf.float32,
        initializer=tf.contrib.layers.xavier_initializer_conv2d())
        # 卷积处理
        conv = tf.nn.conv2d(x, kernel, (1, dh, dw, 1), padding='SAME')
        # biases：初始化
        bias_init = tf.constant(0.0, shape=[n_out], dtype=tf.float32)
        biases = tf.Variable(bias_init, trainable=True, name='b')
        # conv + biases
        z = tf.nn.bias_add(conv, biases)
        activation = tf.nn.relu(z, name=scope)
        # 将本层参数kernel和biases存入参数列表
        p += [kernel, biases]
        return activation

# 全连接层
def fcLayer(x, name, n_out, p):
    # 输入x的通道数
    n_in = x.get_shape()[-1].value
    # 设置scope
    with tf.name_scope(name) as scope:
        #
        kernel = tf.get_variable(scope+'w',
        shape=[n_in, n_out], dtype=tf.float32,
        initializer=tf.contrib.layers.xavier_initializer())
        # 将biases初始化一个较小的值以避免dead neuron
        biases = tf.Variable(tf.constant(0.1, shape=[n_out], dtype=tf.float32), name='b')
        # Relu(x * kernel + biases).
        activation = tf.nn.relu_layer(x, kernel, biases, name='b')
        # 将本层参数存入参数列表
        p += [kernel, biases]
        return activation


#
def VGGNet(x, keep_prob):
    p = []
    # 卷积1　卷积核3*3,数量64,步长1*1
    # 输入224*224*3,输出224*224*64
    conv1_1 = convLayer(x, name='conv1_1', kh=3, kw=3, n_out=64, dh=1, dw=1, p=p)
    # 输入224*224*64,输出224*224*64
    conv1_2 = convLayer(conv1_1, name='conv1_2', kh=3, kw=3, n_out=64, dh=1, dw=1, p=p)
    # 池化尺寸2*2,步长2*2
    # 输入224*224*64,输出112*112*64
    pool1 = tf.nn.max_pool(conv1_2,ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],padding='SAME', name='pool1')

    # 卷积2　卷积核3*3,数量128,步长1*1
    # 输入112*112*64,输出112*112*128
    conv2_1 = convLayer(pool1, name='conv2_1', kh=3, kw=3, n_out=128, dh=1, dw=1, p=p)
    # 输入112*112*128,输出112*112*128
    conv2_2 = convLayer(conv2_1, name='conv2_2', kh=3, kw=3, n_out=128, dh=1, dw=1, p=p)
    # 池化尺寸2*2,步长2*2
    # 输入112*112*128,输出56*56*128
    pool2 = tf.nn.max_pool(conv2_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool2')

    # 卷积3　卷积核3*3,数量256,步长1*1
    # 输入56*56*128,输出56*56*256
    conv3_1 = convLayer(pool2, name='conv3_1', kh=3, kw=3, n_out=256, dh=1, dw=1, p=p)
    # 输入56*56*256,输出56*56*256
    conv3_2 = convLayer(conv3_1, name='conv3_2', kh=3, kw=3, n_out=256, dh=1, dw=1, p=p)
    # 输入56*56*256,输出56*56*256
    conv3_3 = convLayer(conv3_2, name='conv3_3', kh=3, kw=3, n_out=256, dh=1, dw=1, p=p)
    # 池化尺寸2*2,步长2*2
    # 输入56*56*256,输出28*28*256
    pool3 = tf.nn.max_pool(conv3_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool3')

    # 卷积4　卷积核3*3,数量512,步长1*1
    # 输入28*28*256,输出28*28*512
    conv4_1 = convLayer(pool3, name='conv4_1', kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
    # 输入28*28*512,输出28*28*512
    conv4_2 = convLayer(conv4_1, name='conv4_2', kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
    # 输入28*28*512,输出28*28*512
    conv4_3 = convLayer(conv4_2, name='conv4_3', kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
    # 池化尺寸2*2,步长2*2
    # 输入28*28*512,输出14*14*512
    pool4 = tf.nn.max_pool(conv4_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool4')

    # 卷积5　卷积核3*3,数量512,步长1*1
    # 输入14*14*512,输出14*14*512
    conv5_1 = convLayer(pool4, name='conv5_1', kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
    # 输入14*14*512,输出14*14*512
    conv5_2 = convLayer(conv5_1, name='conv5_2', kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
    # 输入14*14*512,输出14*14*512
    conv5_3 = convLayer(conv5_2, name='conv5_3', kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
    # 池化尺寸2*2,步长2*2
    # 输入14*14*512,输出7*7*512
    pool5 = tf.nn.max_pool(conv5_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool5')

    shp = pool5.get_shape()
    flattened_shape = shp[1].value * shp[2].value * shp[3].value
    # 将每个样本化为长度为7*7*512的一维向量
    resh1 = tf.reshape(pool5, [-1, flattened_shape], name='resh1')

    # 全连接层　FC-4096
    fc6 = fcLayer(resh1, name='fc6', n_out=4096, p=p)
    # Dropout层训练时节点保留率为0.5
    fc6_drop = tf.nn.dropout(fc6, keep_prob, name='fc6_drop')
    # 全连接层　FC-4096
    fc7_drop = fcLayer(fc6_drop, name='fc7', n_out=4096, p=p)
    # 全连接层　FC-1000
    fc8 = fcLayer(fc7_drop, name='fc8', n_out=1000, p=p)

    softmax = tf.nn.softmax(fc8)
    # 求出概率最大的类别
    predictions = tf.argmax(softmax, 1)
    return predictions, softmax, fc8, p


#
def time_tensorflow_run(session, target, feed, info_string):
    num_batches = 100
    num_steps_burn_in = 10
    total_duration = 0.0
    total_duration_squared = 0.0
    for i in range(num_batches + num_steps_burn_in):
        start_time = time.time()
        _ = session.run(target, feed_dict=feed)
        # 持续时间
        duration = time.time() - start_time
        if i >= num_steps_burn_in:
            if not i % 10:
                print '%s: step %d, duration = %.3f' % (datetime.now(), i - num_steps_burn_in, duration)
                # 总持续时间
                total_duration += duration
                # 总持续时间平方和
                total_duration_squared += duration * duration
    # 计算每轮迭代的平均耗时mn,和标准差sd
    mn = total_duration / num_batches
    vr = total_duration_squared / num_batches - mn * mn
    sd = math.sqrt(vr)
    # 打印出每轮迭代耗时
    print '%s: %s across %d steps, %.3f +/- %.3f sec /batch' % (datetime.now(), info_string, num_batches, mn, sd)



#
def run_benchmark():
    with tf.Graph().as_default():
        batch_size = 32
        image_size = 224
        # 生成尺寸为224*224的随机图片
        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)
        # 构建VGGNet-16网络
        predictions, softmax, fc8, p =VGGNet(images, keep_prob)
        # 创建Session并初始化全局参数
        init = tf.global_variables_initializer()
        sess = tf.Session()
        sess.run(init)
        # 将keep_prob置为1,进行预测predictions,测评forward运算时间
        time_tensorflow_run(sess, predictions, {keep_prob:1.0}, 'Froward')
        objective = tf.nn.l2_loss(fc8)
        grad = tf.gradients(objective, p)
        # 将keep_prob置为0.5,求解梯度操作grad,测评backward运算时间
        time_tensorflow_run(sess, grad, {keep_prob:0.5}, 'Forward-backward')


if __name__ == '__main__':
    run_benchmark()

打印输出：

2017-11-24 09:47:57.623426: step 0, duration = 90.848
2017-11-24 10:00:07.691008: step 10, duration = 61.761
2017-11-24 10:10:50.857570: step 20, duration = 66.089
2017-11-24 10:22:25.293988: step 30, duration = 73.029
2017-11-24 10:34:29.140638: step 40, duration = 70.612
2017-11-24 10:45:46.713019: step 50, duration = 69.431
2017-11-24 10:57:05.660523: step 60, duration = 70.731
2017-11-24 11:08:44.128224: step 70, duration = 71.724
2017-11-24 11:20:48.707626: step 80, duration = 76.572
2017-11-24 11:32:42.663774: step 90, duration = 68.053
2017-11-24 11:42:23.652837: Froward across 100 steps, 7.188 +/- 21.691 sec /batch
2017-11-24 12:08:49.406644: step 0, duration = 101.215
2017-11-24 12:32:49.628456: step 10, duration = 170.579
2017-11-24 12:56:07.453533: step 20, duration = 140.820
2017-11-24 13:34:04.651331: step 30, duration = 208.970
2017-11-24 14:07:06.223686: step 40, duration = 212.811
2017-11-24 14:42:08.806884: step 50, duration = 203.940
2017-11-24 15:16:49.314282: step 60, duration = 206.452
2017-11-24 15:51:43.567262: step 70, duration = 189.478
2017-11-24 16:28:05.278715: step 80, duration = 224.151
2017-11-24 17:00:15.769221: step 90, duration = 184.321
2017-11-24 17:29:12.099079: Forward-backward across 100 steps, 18.427 +/- 56.436 sec /batch

利用VGGNet-16图像识别：

利用VGGNet-16图像识别并输出识别结果.

import tensorflow as tf
import numpy as np
from scipy.misc import imread, imresize
from imagenet_classes import class_names
import cv2

# 卷积层
def convLayer(x, name, kh, kw, n_out, dh, dw, p):
    # 输入数据的通道数
    n_in = x.get_shape()[-1].value
    # 设置scope
    with tf.name_scope(name) as scope:
        # 卷积核参数
        kernel = tf.Variable(tf.truncated_normal([kh, kw, n_in, n_out], dtype=tf.float32, stddev=1e-1), name='weights')
        # 卷积处理
        conv = tf.nn.conv2d(x, kernel, (1, dh, dw, 1), padding='SAME')
        # biases：初始化
        bias_init = tf.constant(0.0, shape=[n_out], dtype=tf.float32)
        biases = tf.Variable(bias_init, trainable=True, name='biases')
        # conv + biases
        z = tf.nn.bias_add(conv, biases)
        activation = tf.nn.relu(z, name=scope)
        # 将本层参数kernel和biases存入参数列表
        p += [kernel, biases]
        return activation, p


class VGGNet16:
    def __init__(self, imgs, weights=None, sess=None):
        self.imgs = imgs
        self.convlayers()
        self.fc_layers()
        self.probs = tf.nn.softmax(self.fc3l)
        if weights is not None and sess is not None:
            self.load_weights(weights, sess)

    def convlayers(self):
        self.parameters = []
        # 图像预处理：将RGB图像的像素值的范围设置为0-255,然后减去平均图像值
        with tf.name_scope('preprocess') as scope:
            mean = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32, shape=[1, 1, 1, 3], name='img_mean')
            images = self.imgs-mean
        # conv1_1
        self.conv1_1, self.parameters =convLayer(images, name='conv1_1',
                                    kh=3, kw=3, n_out=64, dh=1, dw=1, p=self.parameters)
        # conv1_2
        self.conv1_2, self.parameters =convLayer(self.conv1_1, name='conv1_2',
                                    kh=3, kw=3, n_out=64, dh=1, dw=1, p=self.parameters)
        # pool1
        self.pool1 = tf.nn.max_pool(self.conv1_2, ksize=[1, 2, 2, 1],
                               strides=[1, 2, 2, 1], padding='SAME', name='pool1')
        # conv2_1
        self.conv2_1, self.parameters =convLayer(self.pool1, name='conv2_1',
                                    kh=3, kw=3, n_out=128, dh=1, dw=1, p=self.parameters)
        # conv2_2
        self.conv2_2, self.parameters =convLayer(self.conv2_1, name='conv2_2',
                                    kh=3, kw=3, n_out=128, dh=1, dw=1, p=self.parameters)
        # pool2
        self.pool2 = tf.nn.max_pool(self.conv2_2, ksize=[1, 2, 2, 1],
                               strides=[1, 2, 2, 1], padding='SAME', name='pool2')
        # conv3_1
        self.conv3_1, self.parameters =convLayer(self.pool2, name='conv3_1',
                                    kh=3, kw=3, n_out=256, dh=1, dw=1, p=self.parameters)
        # conv3_2
        self.conv3_2, self.parameters =convLayer(self.conv3_1, name='conv3_2',
                                    kh=3, kw=3, n_out=256, dh=1, dw=1, p=self.parameters)
        # conv3_3
        self.conv3_3, self.parameters =convLayer(self.conv3_2, name='conv3_3',
                                    kh=3, kw=3, n_out=256, dh=1, dw=1, p=self.parameters)
        # pool3
        self.pool3 = tf.nn.max_pool(self.conv3_3, ksize=[1, 2, 2, 1],
                                    strides=[1, 2, 2, 1],padding='SAME', name='pool3')
        # conv4_1
        self.conv4_1, self.parameters =convLayer(self.pool3, name='conv4_1',
                                    kh=3, kw=3, n_out=512, dh=1, dw=1, p=self.parameters)
        # conv4_2
        self.conv4_2, self.parameters =convLayer(self.conv4_1, name='conv4_2',
                                    kh=3, kw=3, n_out=512, dh=1, dw=1, p=self.parameters)
        # conv4_3
        self.conv4_3, self.parameters =convLayer(self.conv4_2, name='conv4_3',
                                    kh=3, kw=3, n_out=512, dh=1, dw=1, p=self.parameters)
        # pool4
        self.pool4 = tf.nn.max_pool(self.conv4_3, ksize=[1, 2, 2, 1],
                               strides=[1, 2, 2, 1], padding='SAME', name='pool4')
        # conv5_1
        self.conv5_1, self.parameters =convLayer(self.pool4, name='conv5_1',
                                    kh=3, kw=3, n_out=512, dh=1, dw=1, p=self.parameters)
        # conv5_2
        self.conv5_2, self.parameters =convLayer(self.conv5_1, name='conv5_2',
                                    kh=3, kw=3, n_out=512, dh=1, dw=1, p=self.parameters)
        # conv5_3
        self.conv5_3, self.parameters =convLayer(self.conv5_2, name='conv5_3',
                                    kh=3, kw=3, n_out=512, dh=1, dw=1, p=self.parameters)
        # pool5
        self.pool5 = tf.nn.max_pool(self.conv5_3, ksize=[1, 2, 2, 1],
                               strides=[1, 2, 2, 1],padding='SAME',name='pool4')

    def fc_layers(self):
        # 全连接层fc1
        with tf.name_scope('fc1') as scope:
            shape = int(np.prod(self.pool5.get_shape()[1:]))
            fc1w = tf.Variable(tf.truncated_normal([shape, 4096], dtype=tf.float32,stddev=1e-1), name='weights')
            fc1b = tf.Variable(tf.constant(1.0, shape=[4096], dtype=tf.float32), trainable=True, name='biases')
            pool5_flat = tf.reshape(self.pool5, [-1, shape])
            fc1l = tf.nn.bias_add(tf.matmul(pool5_flat, fc1w), fc1b)
            self.fc1 = tf.nn.relu(fc1l)
            self.parameters += [fc1w, fc1b]

        # 全连接层fc2
        with tf.name_scope('fc2') as scope:
            fc2w = tf.Variable(tf.truncated_normal([4096, 4096], dtype=tf.float32, stddev=1e-1), name='weights')
            fc2b = tf.Variable(tf.constant(1.0, shape=[4096], dtype=tf.float32), trainable=True, name='biases')
            fc2l = tf.nn.bias_add(tf.matmul(self.fc1, fc2w), fc2b)
            self.fc2 = tf.nn.relu(fc2l)
            self.parameters += [fc2w, fc2b]

        # 全连接层fc3
        with tf.name_scope('fc3') as scope:
            fc3w = tf.Variable(tf.truncated_normal([4096, 1000], dtype=tf.float32, stddev=1e-1), name='weights')
            fc3b = tf.Variable(tf.constant(1.0, shape=[1000], dtype=tf.float32), trainable=True, name='biases')
            self.fc3l = tf.nn.bias_add(tf.matmul(self.fc2, fc3w), fc3b)
            self.parameters += [fc3w, fc3b]
    # 加载训练好的文件模型
    def load_weights(self, weight_file, sess):
        weights = np.load(weight_file)
        keys = sorted(weights.keys())
        for i, k in enumerate(keys):
            print i, k, np.shape(weights[k])
            sess.run(self.parameters[i].assign(weights[k]))

if __name__ == '__main__':
    sess = tf.Session()
    imgs = tf.placeholder(tf.float32, [None, 224, 224, 3])
    # VGGNet16模型
    vgg = VGGNet16(imgs, 'vgg16_weights.npz', sess)
    # img = cv2.imread('cat.jpg')
    img = cv2.imread('crane.jpg')
    img1 = imresize(img, (224, 224))
    prob = sess.run(vgg.probs, feed_dict={vgg.imgs: [img1]})[0]
    # 输出概率最高的前5种类别,以及对应的概率大小
    preds = (np.argsort(prob)[::-1])[0:5]
    for p in preds:
        print class_names[p], prob[p]
    # 概率最大的类
    res = class_names[preds[0]]
    font = cv2.FONT_HERSHEY_SIMPLEX
    # 显示类的名字
    cv2.putText(img, res, (int(img.shape[0] / 10), int(img.shape[1] / 10)), font, 1, (255, 0, 0), 2)
    cv2.imshow("test", img)
    cv2.waitKey(0)

输出：

0 conv1_1_W (3, 3, 3, 64)
1 conv1_1_b (64,)
2 conv1_2_W (3, 3, 64, 64)
3 conv1_2_b (64,)
4 conv2_1_W (3, 3, 64, 128)
5 conv2_1_b (128,)
6 conv2_2_W (3, 3, 128, 128)
7 conv2_2_b (128,)
8 conv3_1_W (3, 3, 128, 256)
9 conv3_1_b (256,)
10 conv3_2_W (3, 3, 256, 256)
11 conv3_2_b (256,)
12 conv3_3_W (3, 3, 256, 256)
13 conv3_3_b (256,)
14 conv4_1_W (3, 3, 256, 512)
15 conv4_1_b (512,)
16 conv4_2_W (3, 3, 512, 512)
17 conv4_2_b (512,)
18 conv4_3_W (3, 3, 512, 512)
19 conv4_3_b (512,)
20 conv5_1_W (3, 3, 512, 512)
21 conv5_1_b (512,)
22 conv5_2_W (3, 3, 512, 512)
23 conv5_2_b (512,)
24 conv5_3_W (3, 3, 512, 512)
25 conv5_3_b (512,)
26 fc6_W (25088, 4096)
27 fc6_b (4096,)
28 fc7_W (4096, 4096)
29 fc7_b (4096,)
30 fc8_W (4096, 1000)
31 fc8_b (1000,)
crane 0.977137
red-breasted merganser, Mergus serrator 0.00527312
peacock 0.0034342
black stork, Ciconia nigra 0.00178344
magpie 0.00142922

---

VGGNet相关连接：

1.牛津大学计算机视觉组Visual Geometry Group

2.Very Deep Convolutional Network For Large-scale Image Recognition论文连接

3.ImageNet-2014大规模视觉识别挑战赛Large Scale Visual Recognition Challenge 2014 (ILSVRC2014)
3.TensorFlow实现VGG