很多深度神經網路模型需要載入預訓練過的Vgg引數，比如說：風格遷移、目標檢測、影象標註等計算機視覺中常見的任務。那麼到底如何載入Vgg模型呢？Vgg檔案的引數到底有何意義呢？載入後的模型該如何使用呢？

本文將以Vgg19為例子，詳細說明Tensorflow如何載入Vgg預訓練模型。

實驗環境

GTX1050-ti,cuda9.0

Window10,Tensorflow 1.12

展示Vgg19構造

import tensorflow as tf
 
import numpy as np
import scipy.io
 
data_path = 'model/vgg19.mat' # data_path指下載下來的Vgg19預訓練模型的檔案地址
 
# 讀取Vgg19檔案
data = scipy.io.loadmat(data_path)
# 列印Vgg19的資料型別及其組成
print("type: ",type(data))
print("data.keys: ",data.keys())
 
# 得到對應卷積核的矩陣
weights = data['layers'][0]
# 定義Vgg19的組成
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'
)
 
# 列印Vgg19不同卷積層所對應的維度
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)
 
 
程式碼輸出結果如下：
type: <class 'dict'>
data.keys: dict_keys(['__header__','__version__','__globals__','layers','meta'])
 
conv1_1: (3,3,64)
relu1_1
conv1_2: (3,64,64)
relu1_2
pool1
conv2_1: (3,128)
relu2_1
conv2_2: (3,128,128)
relu2_2
pool2
conv3_1: (3,256)
relu3_1
conv3_2: (3,256,256)
relu3_2
conv3_3: (3,256)
relu3_3
conv3_4: (3,256)
relu3_4
pool3
conv4_1: (3,512)
relu4_1
conv4_2: (3,512,512)
relu4_2
conv4_3: (3,512)
relu4_3
conv4_4: (3,512)
relu4_4
pool4
conv5_1: (3,512)
relu5_1
conv5_2: (3,512)
relu5_2
conv5_3: (3,512)
relu5_3
conv5_4: (3,512)
relu5_4

那麼Vgg19真實的網路結構是怎麼樣子的呢，如下圖所示：

在本文，主要討論卷積模組，大家通過對比可以發現，我們打印出來的Vgg19結構及其卷積核的構造的確如論文中給出的Vgg19結構一致。

構建Vgg19模型

def _conv_layer(input,weights,bias):
 conv = tf.nn.conv2d(input,tf.constant(weights),strides=(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,padding='SAME')
 
class VGG19:
 layers = (
  'conv1_1','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):
  # 定義net用來儲存模型每一步輸出的特徵圖
  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,(1,3))
     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)
    # 在每一步都儲存當前輸出的特徵圖
    net[name] = current
 
  return net

在上面的程式碼中，我們定義了一個Vgg19的類別專門用來載入Vgg19模型，並且將每一層卷積得到的特徵圖儲存到net中，最後返回這個net，用於程式碼後續的處理。

測試Vgg19模型

在給出Vgg19的構造模型後，我們下一步就是如何用它，我們的思路如下：

載入本地圖片

定義Vgg19模型，傳入本地圖片

得到返回每一層的特徵圖

image_path = "data/test.jpg" # 本地的測試圖片
 
image_raw = tf.gfile.GFile(image_path,'rb').read()
# 一定要tf.float()，否則會報錯
image_decoded = tf.to_float(tf.image.decode_jpeg(image_raw))
 
# 擴充套件圖片的維度，從三維變成四維，符合Vgg19的輸入介面
image_expand_dim = tf.expand_dims(image_decoded,0)
 
# 定義Vgg19模型
vgg19 = VGG19(data_path)
net = vgg19.feed_forward(image_expand_dim,'vgg19')
print(net)
 
程式碼結果如下所示：
{'conv1_1': <tf.Tensor 'vgg19_1/BiasAdd:0' shape=(1,?,64) dtype=float32>,'relu1_1': <tf.Tensor 'vgg19_1/Relu:0' shape=(1,'conv1_2': <tf.Tensor 'vgg19_1/BiasAdd_1:0' shape=(1,'relu1_2': <tf.Tensor 'vgg19_1/Relu_1:0' shape=(1,'pool1': <tf.Tensor 'vgg19_1/MaxPool:0' shape=(1,'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,'conv2_2': <tf.Tensor 'vgg19_1/BiasAdd_3:0' shape=(1,'relu2_2': <tf.Tensor 'vgg19_1/Relu_3:0' shape=(1,'pool2': <tf.Tensor 'vgg19_1/MaxPool_1:0' shape=(1,'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,'conv3_2': <tf.Tensor 'vgg19_1/BiasAdd_5:0' shape=(1,'relu3_2': <tf.Tensor 'vgg19_1/Relu_5:0' shape=(1,'conv3_3': <tf.Tensor 'vgg19_1/BiasAdd_6:0' shape=(1,'relu3_3': <tf.Tensor 'vgg19_1/Relu_6:0' shape=(1,'conv3_4': <tf.Tensor 'vgg19_1/BiasAdd_7:0' shape=(1,'relu3_4': <tf.Tensor 'vgg19_1/Relu_7:0' shape=(1,'pool3': <tf.Tensor 'vgg19_1/MaxPool_2:0' shape=(1,'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,'conv4_2': <tf.Tensor 'vgg19_1/BiasAdd_9:0' shape=(1,'relu4_2': <tf.Tensor 'vgg19_1/Relu_9:0' shape=(1,'conv4_3': <tf.Tensor 'vgg19_1/BiasAdd_10:0' shape=(1,'relu4_3': <tf.Tensor 'vgg19_1/Relu_10:0' shape=(1,'conv4_4': <tf.Tensor 'vgg19_1/BiasAdd_11:0' shape=(1,'relu4_4': <tf.Tensor 'vgg19_1/Relu_11:0' shape=(1,'pool4': <tf.Tensor 'vgg19_1/MaxPool_3:0' shape=(1,'conv5_1': <tf.Tensor 'vgg19_1/BiasAdd_12:0' shape=(1,'relu5_1': <tf.Tensor 'vgg19_1/Relu_12:0' shape=(1,'conv5_2': <tf.Tensor 'vgg19_1/BiasAdd_13:0' shape=(1,'relu5_2': <tf.Tensor 'vgg19_1/Relu_13:0' shape=(1,'conv5_3': <tf.Tensor 'vgg19_1/BiasAdd_14:0' shape=(1,'relu5_3': <tf.Tensor 'vgg19_1/Relu_14:0' shape=(1,'conv5_4': <tf.Tensor 'vgg19_1/BiasAdd_15:0' shape=(1,'relu5_4': <tf.Tensor 'vgg19_1/Relu_15:0' shape=(1,512) dtype=float32>}

本文提供的測試程式碼是完成正確的，已經避免了很多使用Vgg19預訓練模型的坑操作，比如：給圖片新增維度，轉換讀取圖片的的格式等，為什麼這麼做的詳細原因可參考我的另一篇部落格：Tensorflow載入Vgg預訓練模型的幾個注意事項。

到這裡，如何使用tensorflow讀取Vgg19模型結束了，若是大家有其他疑惑，可在評論區留言，會定時回答。

好了，以上就是小編分享給大家的全部內容了，希望能給大家一個參考，也希望大家多多支援我們。