1、AlexNet网络模型，pytorch1.1.0 实现 　　

　　注意：AlexNet，in_img_size >=64 输入图片矩阵的大小要大于等于64

# coding:utf-8
import torch.nn as nn
import torch

class alex_net(nn.Module):
    def __init__(self,in_img_rgb=3,in_img_size=64,out_class=1000,in_fc_size=9216):
        super(alex_net,self).__init__()

        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels=in_img_rgb, out_channels=in_img_size, kernel_size=11, stride=4, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(in_channels=in_img_size,out_channels=192,kernel_size=5,stride=1,padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
        )
        self.conv3 = nn.Sequential(
            nn.Conv2d(in_channels=192,out_channels=384,kernel_size=3,stride=1,padding=1),
            nn.ReLU()
        )
        self.conv4 = nn.Sequential(
            nn.Conv2d(in_channels=384,out_channels=256,kernel_size=3,stride=1,padding=1),
            nn.ReLU()
        )
        self.conv5 = nn.Sequential(
            nn.Conv2d(in_channels=256,out_channels=256,kernel_size=3,stride=1,padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
        )

        self.fc1 = nn.Sequential(
            nn.Dropout(p=0.5),
            nn.Linear(in_features=in_fc_size, out_features=4096, bias=True),
            nn.ReLU(),
            nn.Dropout(p=0.5)
        )

        self.fc2 = nn.Sequential(
            nn.Linear(in_features=4096, out_features=4096, bias=True),
            nn.ReLU()
        )
        self.fc3 = nn.Sequential(
            nn.Linear(in_features=4096, out_features=out_class, bias=True)
        )

        self.conv_list = [self.conv1,self.conv2,self.conv3,self.conv4,self.conv5]

        self.fc_list = [self.fc1,self.fc2,self.fc3]

    def forward(self, x):

        for conv in self.conv_list:
            x = conv(x)

        fc = x.view(x.size(0), -1)

        # 查看全连接层的参数：in_fc_size  的值
        # print("alexnet_model_fc:",fc.size(1))

        for fc_item in self.fc_list:
            fc = fc_item(fc)

        return fc


# 检测 gpu是否可用
CUDA = torch.cuda.is_available()

print(CUDA)
if CUDA:
    alex_net_model = alex_net(in_img_rgb=1, in_img_size=80, out_class=13,in_fc_size=256).cuda()
else:
    alex_net_model = alex_net(in_img_rgb=1, in_img_size=80, out_class=13,in_fc_size=256)

print(alex_net_model)

# 优化方法
optimizer = torch.optim.Adam(alex_net_model.parameters())
# 损失函数
loss_func = nn.MultiLabelSoftMarginLoss()#nn.CrossEntropyLoss()

# 批次训练分割数据集
def batch_training_data(x_train,y_train,batch_size,i):
    n = len(x_train)
    left_limit = batch_size*i
    right_limit = left_limit+batch_size
    if n>=right_limit:
        return x_train[left_limit:right_limit,:,:,:],y_train[left_limit:right_limit,:]
    else:
        return x_train[left_limit:, :, :, :], y_train[left_limit:, :]

2、训练网络，自定义数据集

#  coding:utf-8
import time
import os
import torch
import numpy as np
from data_processing import get_DS
# from CNN_nework_model import cnn_face_discern_model
from torch.autograd import Variable
from alexnet_model import optimizer, alex_net_model, loss_func, batch_training_data,CUDA
from sklearn.metrics import accuracy_score

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

st = time.time()
# 获取训练集与测试集以 8:2 分割

img_resize = 80
x_,y_,y_true,label = get_DS(img_resize)

label_number = len(label)

x_train,y_train = x_[:960,:,:,:].reshape((960,1,img_resize,img_resize)),y_[:960,:]

x_test,y_test = x_[1250:,:,:,:].reshape((50,1,img_resize,img_resize)),y_[1250:,:]

y_test_label = y_true[1250:]

print(time.time() - st)
print(x_train.shape,x_test.shape)

batch_size = 128
n = int(len(x_train)/batch_size)+1



for epoch in range(100):
    global loss
    for batch in range(n):
        x_training,y_training = batch_training_data(x_train,y_train,batch_size,batch)
        batch_x,batch_y = Variable(torch.from_numpy(x_training)).float(),Variable(torch.from_numpy(y_training)).float()
        if CUDA:
            batch_x=batch_x.cuda()
            batch_y=batch_y.cuda()

        out = alex_net_model(batch_x)
        loss = loss_func(out, batch_y)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    # 测试精准度
    if epoch%5==0:
        global x_test_tst
        if CUDA:
            x_test_tst = Variable(torch.from_numpy(x_test)).float().cuda()
        y_pred = alex_net_model(x_test_tst)

        y_predict = np.argmax(y_pred.cpu().data.numpy(),axis=1)
        # print(y_test_label,"\n",y_predict)
        acc = accuracy_score(y_test_label,y_predict)

        print("loss={} aucc={}".format(loss.cpu().data.numpy(),acc))

# 保存模型
# torch.save(model.state_dict(),'save_torch_model/face_image_recognition_model.pkl')

# 导入模型
# model.load_state_dict(torch.load('params.pkl'))

# 两种保存模型的方法
# https://blog.csdn.net/u012436149/article/details/68948816/

3、训练输出日志

True
alex_net(
  (conv1): Sequential(
    (0): Conv2d(1, 80, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))
    (1): ReLU()
    (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (conv2): Sequential(
    (0): Conv2d(80, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (1): ReLU()
    (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (conv3): Sequential(
    (0): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU()
  )
  (conv4): Sequential(
    (0): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU()
  )
  (conv5): Sequential(
    (0): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU()
    (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (fc1): Sequential(
    (0): Dropout(p=0.5)
    (1): Linear(in_features=256, out_features=4096, bias=True)
    (2): ReLU()
    (3): Dropout(p=0.5)
  )
  (fc2): Sequential(
    (0): Linear(in_features=4096, out_features=4096, bias=True)
    (1): ReLU()
  )
  (fc3): Sequential(
    (0): Linear(in_features=4096, out_features=13, bias=True)
  )
)
0.8886234760284424
(960, 1, 80, 80) (50, 1, 80, 80)
loss=0.3137727379798889 aucc=0.02
loss=0.2404210865497589 aucc=0.08
loss=0.18966872990131378 aucc=0.16
loss=0.10794774442911148 aucc=0.42
loss=0.13021017611026764 aucc=0.78
loss=0.012793565168976784 aucc=0.84
loss=0.01140566635876894 aucc=0.9
loss=0.0007940902141854167 aucc=0.88
loss=0.0029846576508134604 aucc=0.92
loss=0.007708669640123844 aucc=0.92
loss=0.00024750467855483294 aucc=0.96
loss=0.0004877769388258457 aucc=0.94
loss=0.009000929072499275 aucc=0.92
loss=0.005286205094307661 aucc=0.86
loss=5.5937391152838245e-05 aucc=0.92
loss=0.002650830429047346 aucc=0.92
loss=0.003015386639162898 aucc=0.94
loss=8.692526171216741e-05 aucc=0.94
loss=0.0021193104330450296 aucc=0.96
loss=7.769006333546713e-05 aucc=0.94