Tensorflow加载模型实现图像分类识别流程详解

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import os
import scipy.io
import scipy.misc
from imagenet_classes import class_names

def _conv_layer(input,weight,bias):
    conv = tf.nn.conv2d(input,weight,strides=[1,1,1,1],padding="SAME")
    return tf.nn.bias_add(conv,bias)
def _pool_layer(input):
    return tf.nn.max_pool(input,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME")
def preprocess(image,mean_pixel):
    '''简单预处理,全部图片减去平均值'''
    return image-mean_pixel
def unprocess(image,mean_pixel):
    return image+mean_pixel

def imread(path):
    return scipy.misc.imread(path)
def imsave(image,path):
    img = np.clip(image,0,255).astype(np.int8)
    scipy.misc.imsave(path,image)

def net(data_path,input_image,sess=None):
    """
    读取VGG模型参数,搭建VGG网络
    :param data_path: VGG模型文件位置
    :param input_image: 输入测试图像
    :return:
    """
    layers = (
        'conv1_1', 'conv1_2', 'pool1',
        'conv2_1', 'conv2_2', 'pool2',
        'conv3_1', 'conv3_2', 'conv3_3','conv3_4', 'pool3',
        'conv4_1', 'conv4_2', 'conv4_3','conv4_4', 'pool4',
        'conv5_1', 'conv5_2', 'conv5_3','conv5_4', 'pool5',
          'fc1'  ,   'fc2'  ,   'fc3'  ,
        'softmax'
    )
    data = scipy.io.loadmat(data_path)
    mean = data["normalization"][0][0][0][0][0]
    input_image = np.array([preprocess(input_image, mean)]).astype(np.float32)#去除平均值
    net = {}
    current = input_image
    net["src_image"] = tf.constant(current)  # 存储数据
    count = 0 #计数存储
    for i in range(43):
        if str(data['layers'][0][i][0][0][0][0])[:4] == ("relu"):
            continue
        if str(data['layers'][0][i][0][0][0][0])[:4] == ("pool"):
            current = _pool_layer(current)
        elif str(data['layers'][0][i][0][0][0][0]) == ("softmax"):
            current = tf.nn.softmax(current)
        elif i == (37):
            shape = int(np.prod(current.get_shape()[1:]))
            current = tf.reshape(current, [-1, shape])
            kernels, bias = data['layers'][0][i][0][0][0][0]
            kernels = np.reshape(kernels,[-1,4096])
            bias = bias.reshape(-1)
            current = tf.nn.relu(tf.add(tf.matmul(current,kernels),bias))
        elif i == (39):
            kernels, bias = data['layers'][0][i][0][0][0][0]
            kernels = np.reshape(kernels,[4096,4096])
            bias = bias.reshape(-1)
            current = tf.nn.relu(tf.add(tf.matmul(current,kernels),bias))
        elif i == 41:
            kernels, bias = data['layers'][0][i][0][0][0][0]
            kernels = np.reshape(kernels, [4096, 1000])
            bias = bias.reshape(-1)
            current = tf.add(tf.matmul(current, kernels), bias)
        else:
            kernels,bias = data['layers'][0][i][0][0][0][0]
            #注意VGG存储方式为[,]
            #kernels = np.transpose(kernels,[1,0,2,3])
            bias = bias.reshape(-1)#降低维度
            current = tf.nn.relu(_conv_layer(current,kernels,bias))
        net[layers[count]] = current #存储数据
        count += 1
    return net, mean

if __name__ == '__main__':
    VGG_PATH = "./one/imagenet-vgg-verydeep-19.mat"
    IMG_PATH = './one/3.jpg'
    input_image =imread(IMG_PATH)
    shape = (1, input_image.shape[0], input_image.shape[1], input_image.shape[2])
    with tf.Session() as sess:
        image = tf.placeholder('float', shape=shape)
        nets, mean_pixel, all_layers= net(VGG_PATH, image)
        input_image_pre=np.array([preprocess(input_image,mean_pixel)])
        layers = all_layers
        for i , layer in enumerate(layers):
            print("[%d/%d] %s" % (i+1,len(layers),layers))
            features = nets[layer].eval(feed_dict={image:input_image_pre})
            print("Type of 'feature' is ",type(features))
            print("Shape of 'features' is  %s" % (features.shape,))
            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()