很多深度神经网络模型需要加载预训练过的Vgg参数，比如说：风格迁移、目标检测、图像标注等计算机视觉中常见的任务。那么到底如何加载Vgg模型呢？Vgg文件的参数到底有何意义呢？加载后的模型该如何使用呢？本文将以Vgg19为例子，详细说明Tensorflow如何加载Vgg预训练模型。

实验环境

• GTX1050-ti,  cuda9.0
• Window10,  Tensorflow 1.12

展示Vgg19构造

import tensorflow as tf

import numpy as np
import scipy.io

data_path = 'model/vgg19.mat'  # data_path指下载下来的Vgg19预训练模型的文件地址

# 读取Vgg19文件
data = scipy.io.loadmat(data_path)
# 打印Vgg19的数据类型及其组成
print("type: ", type(data))
print("data.keys: ", data.keys())

# 得到对应卷积核的矩阵
weights = data['layers'][0]
# 定义Vgg19的组成
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'
)

# 打印Vgg19不同卷积层所对应的维度
for i, name in enumerate(layers):
    kind = name[:4]
    if kind == 'conv':
        print("%s: %s" % (name, weights[i][0][0][2][0][0].shape))
    elif kind == 'relu':
        print(name)
    elif kind == 'pool':
        print(name)


代码输出结果如下：
type:  <class 'dict'>
data.keys:  dict_keys(['__header__', '__version__', '__globals__', 'layers', 'meta'])

conv1_1: (3, 3, 3, 64)
relu1_1
conv1_2: (3, 3, 64, 64)
relu1_2
pool1
conv2_1: (3, 3, 64, 128)
relu2_1
conv2_2: (3, 3, 128, 128)
relu2_2
pool2
conv3_1: (3, 3, 128, 256)
relu3_1
conv3_2: (3, 3, 256, 256)
relu3_2
conv3_3: (3, 3, 256, 256)
relu3_3
conv3_4: (3, 3, 256, 256)
relu3_4
pool3
conv4_1: (3, 3, 256, 512)
relu4_1
conv4_2: (3, 3, 512, 512)
relu4_2
conv4_3: (3, 3, 512, 512)
relu4_3
conv4_4: (3, 3, 512, 512)
relu4_4
pool4
conv5_1: (3, 3, 512, 512)
relu5_1
conv5_2: (3, 3, 512, 512)
relu5_2
conv5_3: (3, 3, 512, 512)
relu5_3
conv5_4: (3, 3, 512, 512)
relu5_4

    那么Vgg19真实的网络结构是怎么样子的呢，如下图所示：

    在本文，主要讨论卷积模块，大家通过对比可以发现，我们打印出来的Vgg19结构及其卷积核的构造的确如论文中给出的Vgg19结构一致。

构建Vgg19模型

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)

def _pool_layer(input):
    return tf.nn.max_pool(input, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1),
            padding='SAME')

class VGG19:
    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'
    )

    def __init__(self, data_path):
        data = scipy.io.loadmat(data_path)

        self.weights = data['layers'][0]

    def feed_forward(self, input_image, scope=None):
        # 定义net用来保存模型每一步输出的特征图
        net = {}
        current = input_image

        with tf.variable_scope(scope):
            for i, name in enumerate(self.layers):
                kind = name[:4]
                if kind == 'conv':
                    kernels = self.weights[i][0][0][2][0][0]
                    bias = self.weights[i][0][0][2][0][1]

                    kernels = np.transpose(kernels, (1, 0, 2, 3))
                    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)
                # 在每一步都保存当前输出的特征图
                net[name] = current

        return net

    在上面的代码中，我们定义了一个Vgg19的类别专门用来加载Vgg19模型，并且将每一层卷积得到的特征图保存到net中，最后返回这个net，用于代码后续的处理。

测试Vgg19模型

    在给出Vgg19的构造模型后，我们下一步就是如何用它，我们的思路如下：

• 加载本地图片
• 定义Vgg19模型，传入本地图片
• 得到返回每一层的特征图

image_path = "data/test.jpg" # 本地的测试图片

image_raw = tf.gfile.GFile(image_path, 'rb').read()
# 一定要tf.float()，否则会报错
image_decoded = tf.to_float(tf.image.decode_jpeg(image_raw))

# 扩展图片的维度，从三维变成四维，符合Vgg19的输入接口
image_expand_dim = tf.expand_dims(image_decoded, 0)

# 定义Vgg19模型
vgg19 = VGG19(data_path)
net = vgg19.feed_forward(image_expand_dim, 'vgg19')
print(net)

代码结果如下所示：
{'conv1_1': <tf.Tensor 'vgg19_1/BiasAdd:0' shape=(1, ?, ?, 64) dtype=float32>,
 'relu1_1': <tf.Tensor 'vgg19_1/Relu:0' shape=(1, ?, ?, 64) dtype=float32>,
 'conv1_2': <tf.Tensor 'vgg19_1/BiasAdd_1:0' shape=(1, ?, ?, 64) dtype=float32>,
 'relu1_2': <tf.Tensor 'vgg19_1/Relu_1:0' shape=(1, ?, ?, 64) dtype=float32>,
 'pool1': <tf.Tensor 'vgg19_1/MaxPool:0' shape=(1, ?, ?, 64) dtype=float32>,
 'conv2_1': <tf.Tensor 'vgg19_1/BiasAdd_2:0' shape=(1, ?, ?, 128) dtype=float32>,
 'relu2_1': <tf.Tensor 'vgg19_1/Relu_2:0' shape=(1, ?, ?, 128) dtype=float32>,
 'conv2_2': <tf.Tensor 'vgg19_1/BiasAdd_3:0' shape=(1, ?, ?, 128) dtype=float32>,
 'relu2_2': <tf.Tensor 'vgg19_1/Relu_3:0' shape=(1, ?, ?, 128) dtype=float32>,
 'pool2': <tf.Tensor 'vgg19_1/MaxPool_1:0' shape=(1, ?, ?, 128) dtype=float32>,
 'conv3_1': <tf.Tensor 'vgg19_1/BiasAdd_4:0' shape=(1, ?, ?, 256) dtype=float32>,
 'relu3_1': <tf.Tensor 'vgg19_1/Relu_4:0' shape=(1, ?, ?, 256) dtype=float32>,
 'conv3_2': <tf.Tensor 'vgg19_1/BiasAdd_5:0' shape=(1, ?, ?, 256) dtype=float32>,
 'relu3_2': <tf.Tensor 'vgg19_1/Relu_5:0' shape=(1, ?, ?, 256) dtype=float32>,
 'conv3_3': <tf.Tensor 'vgg19_1/BiasAdd_6:0' shape=(1, ?, ?, 256) dtype=float32>,
 'relu3_3': <tf.Tensor 'vgg19_1/Relu_6:0' shape=(1, ?, ?, 256) dtype=float32>,
 'conv3_4': <tf.Tensor 'vgg19_1/BiasAdd_7:0' shape=(1, ?, ?, 256) dtype=float32>,
 'relu3_4': <tf.Tensor 'vgg19_1/Relu_7:0' shape=(1, ?, ?, 256) dtype=float32>,
 'pool3': <tf.Tensor 'vgg19_1/MaxPool_2:0' shape=(1, ?, ?, 256) dtype=float32>,
 'conv4_1': <tf.Tensor 'vgg19_1/BiasAdd_8:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu4_1': <tf.Tensor 'vgg19_1/Relu_8:0' shape=(1, ?, ?, 512) dtype=float32>,
 'conv4_2': <tf.Tensor 'vgg19_1/BiasAdd_9:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu4_2': <tf.Tensor 'vgg19_1/Relu_9:0' shape=(1, ?, ?, 512) dtype=float32>,
 'conv4_3': <tf.Tensor 'vgg19_1/BiasAdd_10:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu4_3': <tf.Tensor 'vgg19_1/Relu_10:0' shape=(1, ?, ?, 512) dtype=float32>,
 'conv4_4': <tf.Tensor 'vgg19_1/BiasAdd_11:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu4_4': <tf.Tensor 'vgg19_1/Relu_11:0' shape=(1, ?, ?, 512) dtype=float32>,
 'pool4': <tf.Tensor 'vgg19_1/MaxPool_3:0' shape=(1, ?, ?, 512) dtype=float32>,
 'conv5_1': <tf.Tensor 'vgg19_1/BiasAdd_12:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu5_1': <tf.Tensor 'vgg19_1/Relu_12:0' shape=(1, ?, ?, 512) dtype=float32>,
 'conv5_2': <tf.Tensor 'vgg19_1/BiasAdd_13:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu5_2': <tf.Tensor 'vgg19_1/Relu_13:0' shape=(1, ?, ?, 512) dtype=float32>,
 'conv5_3': <tf.Tensor 'vgg19_1/BiasAdd_14:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu5_3': <tf.Tensor 'vgg19_1/Relu_14:0' shape=(1, ?, ?, 512) dtype=float32>,
 'conv5_4': <tf.Tensor 'vgg19_1/BiasAdd_15:0' shape=(1, ?, ?, 512) dtype=float32>,
 'relu5_4': <tf.Tensor 'vgg19_1/Relu_15:0' shape=(1, ?, ?, 512) dtype=float32>}

    本文提供的测试代码是完成正确的，已经避免了很多使用Vgg19预训练模型的坑操作，比如：给图片添加维度，转换读取图片的的格式等，为什么这么做的详细原因可参考我的另一篇博客：Tensorflow加载Vgg预训练模型的几个注意事项。

    到这里，如何使用tensorflow读取Vgg19模型结束了，若是大家有其他疑惑，可在评论区留言，会定时回答。

备注：本文为作者原创，转载需注明出处！