一、简介
VGG网络在2014年的 ILSVRC localization and classification 两个问题上分别取得了第一名和第二名。VGG网络非常深，通常有16－19层，如果自己训练网络模型的话很浪费时间和计算资源。因此这里采用一种方法获取VGG19模型的模型数据，从而能够更快速的应用到自己的任务中来，
本文在加载模型数据的同时，还可视化图片在网络传播过程中，每一层的输出特征图。让我们能够更直接的观察网络传播的状况。
运行环境为spyder，Python3.5，tensorflow1.2.1
模型名称为： imagenet-vgg-verydeep-19.mat 大家可以在网上下载。
二、VGG19模型结构
模型的每一层结构如下图所示：
三、代码
#加载VGG19模型并可视化一张图片前向传播的过程中每一层的输出
#引入包
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import scipy.io
import scipy.misc
#定义一些函数
#卷积
def _conv_layer(input, weights, bias):
conv = tf.nn.conv2d(input, tf.constant(weights), strides=(1, 1, 1, 1),
padding='SAME')
return tf.nn.bias_add(conv, bias)
#池化
def _pool_layer(input):
return tf.nn.max_pool(input, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1),
padding='SAME')
#减像素均值操作
def preprocess(image, mean_pixel):
return image - mean_pixel
#加像素均值操作
def unprocess(image, mean_pixel):
return image + mean_pixel
#读
def imread(path):
return scipy.misc.imread(path).astype(np.float)
#保存
def imsave(path, img):
img = np.clip(img, 0, 255).astype(np.uint8)
scipy.misc.imsave(path, img)
print ("Functions for VGG ready")
#定义VGG的网络结构，用来存储网络的权重和偏置参数
def net(data_path, input_image):
#拿到每一层对应的参数
layers = (
'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1',
'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',
'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3',
'relu3_3', 'conv3_4', 'relu3_4', 'pool3',
'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3',
'relu4_3', 'conv4_4', 'relu4_4', 'pool4',
'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3',
'relu5_3', 'conv5_4', 'relu5_4'
)
data = scipy.io.loadmat(data_path)
#原网络在训练的过程中，对每张图片三通道都执行了减均值的操作，这里也要减去均值
mean = data['normalization'][0][0][0]
mean_pixel = np.mean(mean, axis=(0, 1))
#print(mean_pixel)
#取到权重参数W和b,这里运气好的话，可以查到VGG模型中每层的参数含义，查不到的
#话可以打印出weights，然后打印每一层的shape，推出其中每一层代表的含义
weights = data['layers'][0]
#print(weights)
net = {}
current = input_image
#取到w和b
for i, name in enumerate(layers):
#:4的含义是只看每一层的前三个字母，从而进行判断
kind = name[:4]
if kind == 'conv':
kernels, bias = weights[i][0][0][0][0]
# matconvnet: weights are [width, height, in_channels, out_channels]\n",
# tensorflow: weights are [height, width, in_channels, out_channels]\n",
#这里width和height是颠倒的，所以要做一次转置运算
kernels = np.transpose(kernels, (1, 0, 2, 3))
#将bias转换为一个维度
bias = bias.reshape(-1)
current = _conv_layer(current, kernels, bias)
elif kind == 'relu':
current = tf.nn.relu(current)
elif kind == 'pool':
current = _pool_layer(current)
net[name] = current
assert len(net) == len(layers)
return net, mean_pixel, layers
print ("Network for VGG ready")
#cwd = os.getcwd()
#这里用的是绝对路径
VGG_PATH = "F:/mnist/imagenet-vgg-verydeep-19.mat"
#需要可视化的图片路径，这里是一只小猫
IMG_PATH = "D:/VS2015Program/cat.jpg"
input_image = imread(IMG_PATH)
#获取图像shape
shape = (1,input_image.shape[0],input_image.shape[1],input_image.shape[2])
#开始会话
with tf.Session() as sess:
image = tf.placeholder('float', shape=shape)
#调用net函数
nets, mean_pixel, all_layers = net(VGG_PATH, image)
#减均值操作（由于VGG网络图片传入前都做了减均值操作，所以这里也用相同的预处理
input_image_pre = np.array([preprocess(input_image, mean_pixel)])
layers = all_layers # For all layers \n",
# layers = ('relu2_1', 'relu3_1', 'relu4_1')\n",
for i, layer in enumerate(layers):
print ("[%d/%d] %s" % (i+1, len(layers), layer))
features = nets[layer].eval(feed_dict={image: input_image_pre})
print (" Type of 'features' is ", type(features))
print (" Shape of 'features' is %s" % (features.shape,))
# Plot response \n",
#画出每一层
if 1:
plt.figure(i+1, figsize=(10, 5))
plt.matshow(features[0, :, :, 0], cmap=plt.cm.gray, fignum=i+1)
plt.title("" + layer)
plt.colorbar()
plt.show()四、程序运行结果
1、print(weights)的结果：
2、程序运行最终结果：
中间层数太多，这里就不展示了。程序最后两层的可视化结果：
