把vgg-face.mat权重迁移到pytorch模型示例

最近使用pytorch时，需要用到一个预训练好的人脸识别模型提取人脸ID特征，想到很多人都在用用vgg-face，但是vgg-face没有pytorch的模型，于是写个vgg-face.mat转到pytorch模型的代码

#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Thu May 10 10:41:40 2018
@author: hy
"""
import torch
import math
import torch.nn as nn
from torch.autograd import Variable
import numpy as np
from scipy.io import loadmat
import scipy.misc as sm
import matplotlib.pyplot as plt

class vgg16_face(nn.Module):
 def __init__(self,num_classes=2622):
   super(vgg16_face,self).__init__()
   inplace = True
   self.conv1_1 = nn.Conv2d(3,64,kernel_size=(3,3),stride=(1,1),padding=(1,1))
   self.relu1_1 = nn.ReLU(inplace)
   self.conv1_2 = nn.Conv2d(64,64,kernel_size=(3,3),stride=(1,1),padding=(1,1))
   self.relu1_2 = nn.ReLU(inplace)
   self.pool1 = nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2), dilation=(1, 1), ceil_mode=False)

self.conv2_1 = nn.Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu2_1 = nn.ReLU(inplace)
   self.conv2_2 = nn.Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu2_2 = nn.ReLU(inplace)
   self.pool2 = nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2), dilation=(1, 1), ceil_mode=False)

self.conv3_1 = nn.Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu3_1 = nn.ReLU(inplace)
   self.conv3_2 = nn.Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu3_2 = nn.ReLU(inplace)
   self.conv3_3 = nn.Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu3_3 = nn.ReLU(inplace)
   self.pool3 = nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2), dilation=(1, 1), ceil_mode=False)

self.conv4_1 = nn.Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu4_1 = nn.ReLU(inplace)
   self.conv4_2 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu4_2 = nn.ReLU(inplace)
   self.conv4_3 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu4_3 = nn.ReLU(inplace)
   self.pool4 = nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2), dilation=(1, 1), ceil_mode=False)

self.conv5_1 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu5_1 = nn.ReLU(inplace)
   self.conv5_2 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu5_2 = nn.ReLU(inplace)
   self.conv5_3 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
   self.relu5_3 = nn.ReLU(inplace)
   self.pool5 = nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2), dilation=(1, 1), ceil_mode=False)

self.fc6 = nn.Linear(in_features=25088, out_features=4096, bias=True)
   self.relu6 = nn.ReLU(inplace)
   self.drop6 = nn.Dropout(p=0.5)

self.fc7 = nn.Linear(in_features=4096, out_features=4096, bias=True)
   self.relu7 = nn.ReLU(inplace)
   self.drop7 = nn.Dropout(p=0.5)
   self.fc8 = nn.Linear(in_features=4096, out_features=num_classes, bias=True)

self._initialize_weights()
 def forward(self,x):
   out = self.conv1_1(x)
   x_conv1 = out
   out = self.relu1_1(out)
   out = self.conv1_2(out)
   out = self.relu1_2(out)
   out = self.pool1(out)
   x_pool1 = out

out = self.conv2_1(out)
   out = self.relu2_1(out)
   out = self.conv2_2(out)
   out = self.relu2_2(out)
   out = self.pool2(out)
   x_pool2 = out

out = self.conv3_1(out)
   out = self.relu3_1(out)
   out = self.conv3_2(out)
   out = self.relu3_2(out)
   out = self.conv3_3(out)
   out = self.relu3_3(out)
   out = self.pool3(out)
   x_pool3 = out

out = self.conv4_1(out)
   out = self.relu4_1(out)
   out = self.conv4_2(out)
   out = self.relu4_2(out)
   out = self.conv4_3(out)
   out = self.relu4_3(out)
   out = self.pool4(out)
   x_pool4 = out

out = self.conv5_1(out)
   out = self.relu5_1(out)
   out = self.conv5_2(out)
   out = self.relu5_2(out)
   out = self.conv5_3(out)
   out = self.relu5_3(out)
   out = self.pool5(out)
   x_pool5 = out

out = out.view(out.size(0),-1)

out = self.fc6(out)
   out = self.relu6(out)
   out = self.fc7(out)
   out = self.relu7(out)
   out = self.fc8(out)

return out, x_pool1, x_pool2, x_pool3, x_pool4, x_pool5

def _initialize_weights(self):
   for m in self.modules():
     if isinstance(m, nn.Conv2d):
       n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
       m.weight.data.normal_(0, math.sqrt(2. / n))
       if m.bias is not None:
         m.bias.data.zero_()
     elif isinstance(m, nn.BatchNorm2d):
       m.weight.data.fill_(1)
       m.bias.data.zero_()
     elif isinstance(m, nn.Linear):
       m.weight.data.normal_(0, 0.01)
       m.bias.data.zero_()

def copy(vgglayers, dstlayer,idx):
 layer = vgglayers[0][idx]
 kernel, bias = layer[0]['weights'][0][0]
 if idx in [33,35]: # fc7, fc8
   kernel = kernel.squeeze()
   dstlayer.weight.data.copy_(torch.from_numpy(kernel.transpose([1,0]))) # matrix format: axb -> bxa
 elif idx == 31: # fc6
   kernel = kernel.reshape(-1,4096)
   dstlayer.weight.data.copy_(torch.from_numpy(kernel.transpose([1,0]))) # matrix format: axb -> bxa
 else:
   dstlayer.weight.data.copy_(torch.from_numpy(kernel.transpose([3,2,1,0]))) # matrix format: axbxcxd -> dxcxbxa
 dstlayer.bias.data.copy_(torch.from_numpy(bias.reshape(-1)))

def get_vggface(vgg_path):
 """1. define pytorch model"""  
 model = vgg16_face()  

"""2. get pre-trained weights and other params"""    
 #vgg_path = "/home/hy/vgg-face.mat" # download from http://www.vlfeat.org/matconvnet/pretrained/
 vgg_weights = loadmat(vgg_path)
 data = vgg_weights
 meta = data['meta']
 classes = meta['classes']
 class_names = classes[0][0]['description'][0][0]
 normalization = meta['normalization']
 average_image = np.squeeze(normalization[0][0]['averageImage'][0][0][0][0])
 image_size = np.squeeze(normalization[0][0]['imageSize'][0][0])
 layers = data['layers']
 # =============================================================================
 # for idx,layer in enumerate(layers[0]):
 #   name = layer[0]['name'][0][0]
 #   print idx,name
 # """
 # 0 conv1_1
 # 1 relu1_1
 # 2 conv1_2
 # 3 relu1_2
 # 4 pool1
 # 5 conv2_1
 # 6 relu2_1
 # 7 conv2_2
 # 8 relu2_2
 # 9 pool2
 # 10 conv3_1
 # 11 relu3_1
 # 12 conv3_2
 # 13 relu3_2
 # 14 conv3_3
 # 15 relu3_3
 # 16 pool3
 # 17 conv4_1
 # 18 relu4_1
 # 19 conv4_2
 # 20 relu4_2
 # 21 conv4_3
 # 22 relu4_3
 # 23 pool4
 # 24 conv5_1
 # 25 relu5_1
 # 26 conv5_2
 # 27 relu5_2
 # 28 conv5_3
 # 29 relu5_3
 # 30 pool5
 # 31 fc6
 # 32 relu6
 # 33 fc7
 # 34 relu7
 # 35 fc8
 # 36 prob
 # """
 # =============================================================================

"""3. load weights to pytorch model"""  
 copy(layers,model.conv1_1,0)
 copy(layers,model.conv1_2,2)
 copy(layers,model.conv2_1,5)
 copy(layers,model.conv2_2,7)
 copy(layers,model.conv3_1,10)
 copy(layers,model.conv3_2,12)
 copy(layers,model.conv3_3,14)
 copy(layers,model.conv4_1,17)
 copy(layers,model.conv4_2,19)
 copy(layers,model.conv4_3,21)
 copy(layers,model.conv5_1,24)
 copy(layers,model.conv5_2,26)
 copy(layers,model.conv5_3,28)
 copy(layers,model.fc6,31)
 copy(layers,model.fc7,33)
 copy(layers,model.fc8,35)
 return model,class_names,average_image,image_size

if __name__ == '__main__':
 """test"""
 vgg_path = "/home/hy/vgg-face.mat" # download from http://www.vlfeat.org/matconvnet/pretrained/
 model,class_names,average_image,image_size = get_vggface(vgg_path)
 imgpath = "/home/hy/e/avg_face.jpg"
 img = sm.imread(imgpath)
 img = sm.imresize(img,[image_size[0],image_size[1]])
 input_arr = np.float32(img)#-average_image # h,w,c
 x = torch.from_numpy(input_arr.transpose((2,0,1))) # c,h,w
 avg = torch.from_numpy(average_image) #
 avg = avg.view(3,1,1).expand(3,224,224)
 x = x - avg
 x = x.contiguous()
 x = x.view(1, x.size(0), x.size(1), x.size(2))
 x = Variable(x)
 out, x_pool1, x_pool2, x_pool3, x_pool4, x_pool5 = model(x)
#  plt.imshow(x_pool1.data.numpy()[0,45]) # plot

来源：https://blog.csdn.net/ying86615791/article/details/80347761