当前位置:首页 > 技术人生 > 人工智能 > 正文内容

VGG19模型训练+读取

2021-06-14人工智能345

VGG-19模型简单介绍

image.png

image.png

image.png

VGG-19模型文件介绍

这里是重难点,VGG-19模型存储的方式有点复杂

  • 可以通过作者文档说明去查看

  • 可以通过在线调试查看结构,对比模型得出结论

imagenet-vgg-verydeep-19.mat文件下载地址

ImageNet 1000种分类以及排列文件下载地址

分析模型文件

首先我们通过scipy模块(当然你可以用其它方式入opencv / sklearn等)读取scipy.io.loadmat()文件

data = scipy.io.loadmat(data_path)

接下来使用Pycharm在线查看data数据结构:

  • 总共有很多参数,我们只关心我们需要关注的,W和B在哪里就行了

  • 注意这里还有一个mean(平均值),因为VGG使用了图像预处理方式是 input - mean,当然这种处理方式在现在看来不怎么好,但是现在我们用人家的模型,需要遵照人家的意思.

  • 这里主要关注"layers"和"normalization"

image.png

mean值查看

首先查看normalization的mean值:

  • 注意我们使用的是RGB输入,且图像大小为224*224,也就是说图像是[1 * 224 * 224 *3]格式

  • 那么我们的mean肯定就是[X,Y,Z]的格式,因为是RGB的平均值

这里使用笨方法,直接带入:

data["normalization"][0][0][0][0][0]

先看大概看一下数据结构,然后一个一个测试,因为在线调试,这样很快,比你精确的去找要快很多.

image.png

Weight和Bias查看

image.png

这里和mean查看方法不同,W和B得自习查看数据结构,因为太复杂了

  • 从下面的图看到存储的43个参数

  • 注意里面的Relu是没有数据的,因为Relu就是一个函数

  • 注意Pool的参数是固定的,因为大小为:[1,2,2,1],步长[1,2,2,1],这里可以自己写,也可以读取参数

  • Weight Bias是存放在Relu Pool 中间的,而且两个值存在一起的.

  • 后期代码需要进一步处理,后面会说到.

  • 再不明白的,自己画个VGG-19网络就知道了.

image.png

读取代码

mean = data["normalization"][0][0][0][0][0]
data['layers'][0][i][0][0][0][0])

读取模型

这里默认大家已经会CNN的基本操作了,代码不做详细说明

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 = os.getcwd()+"/imagenet-vgg-verydeep-19.mat"
    input_image = scipy.misc.imread("234.jpeg")
    input_image = scipy.misc.imresize(input_image,[224,224,3])
    shape = (1, input_image.shape[0], input_image.shape[1], input_image.shape[2])
    #image = tf.placeholder('float', shape=shape)
    with tf.Session() as sess:
        nets, mean_pixel, = net(VGG_PATH, input_image, sess=sess)
        #print(sess.run(nets,feed_dict={image:input_image}))
        nets = sess.run(nets)
        '''
        for key, values in nets.items():
            if len(values.shape)<4:
                continue
            plt.figure(key)
            plt.matshow(values[0, :, :, 0],)
            plt.title(key)
            plt.colorbar()
            plt.show()
        '''
        #打印概率最大的三个数据
        net_sort    = list(reversed(np.argsort(nets["softmax"]).reshape(-1).tolist()))
        net_softmax = nets["softmax"].reshape(-1).tolist()
        for i in range(3):
            print(class_names[net_sort[i]],": ",net_softmax[net_sort[i]])

输入图片:

image.png

结果显示:

image.png

百度验证:

image.png

训练代码

以下是训练的代码:(这部分随便看看就好,比较简单~~主要看思想和细节!)

########################################################################################
# Davi Frossard, 2016                                                                  #
# VGG16 implementation in TensorFlow                                                   #
# Details:                                                                             #
# http://www.cs.toronto.edu/~frossard/post/vgg16/                                      #
#                                                                                      #
# Model from https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md     #
# Weights from Caffe converted using https://github.com/ethereon/caffe-tensorflow      #
########################################################################################
import tensorflow as tf
import numpy as np
from scipy.misc import imread, imresize
from imagenet_classes import class_names
class vgg16:
    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 = []
        # zero-mean input
        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
        with tf.name_scope('conv1_1') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 3, 64], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(images, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[64], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv1_1 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv1_2
        with tf.name_scope('conv1_2') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 64, 64], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv1_1, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[64], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv1_2 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # 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
        with tf.name_scope('conv2_1') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 64, 128], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.pool1, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[128], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv2_1 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv2_2
        with tf.name_scope('conv2_2') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 128, 128], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv2_1, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[128], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv2_2 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # 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
        with tf.name_scope('conv3_1') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 128, 256], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.pool2, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv3_1 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv3_2
        with tf.name_scope('conv3_2') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 256], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv3_1, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv3_2 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv3_3
        with tf.name_scope('conv3_3') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 256], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv3_2, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv3_3 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # 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
        with tf.name_scope('conv4_1') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 512], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.pool3, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[512], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv4_1 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv4_2
        with tf.name_scope('conv4_2') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 512, 512], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv4_1, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[512], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv4_2 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv4_3
        with tf.name_scope('conv4_3') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 512, 512], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv4_2, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[512], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv4_3 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # 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
        with tf.name_scope('conv5_1') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 512, 512], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.pool4, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[512], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv5_1 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv5_2
        with tf.name_scope('conv5_2') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 512, 512], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv5_1, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[512], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv5_2 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # conv5_3
        with tf.name_scope('conv5_3') as scope:
            kernel = tf.Variable(tf.truncated_normal([3, 3, 512, 512], dtype=tf.float32,
                                                     stddev=1e-1), name='weights')
            conv = tf.nn.conv2d(self.conv5_2, kernel, [1, 1, 1, 1], padding='SAME')
            biases = tf.Variable(tf.constant(0.0, shape=[512], dtype=tf.float32),
                                 trainable=True, name='biases')
            out = tf.nn.bias_add(conv, biases)
            self.conv5_3 = tf.nn.relu(out, name=scope)
            self.parameters += [kernel, biases]
        # 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):
        # 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]
        # 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]
        # 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]
    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])
    vgg = vgg16(imgs, 'vgg16_weights.npz', sess)
    img1 = imread('laska.png', mode='RGB')
    img1 = imresize(img1, (224, 224))
    prob = sess.run(vgg.probs, feed_dict={vgg.imgs: [img1]})[0]
    preds = (np.argsort(prob)[::-1])[0:5]
    for p in preds:
        print class_names[p], prob[p]

扫描二维码推送至手机访问。

本文链接:https://strivers.cn/post/45.html

来源:简书,作者:影醉阏轩窗,链接:https://www.jianshu.com/p/e96b7a9b4229

分享给朋友:
返回列表

上一篇:机器学习的常用算法

没有最新的文章了...

发表评论

看不清,换一张

◎欢迎参与讨论,请在这里发表您的看法和观点。