transfer dense121迁移学习模型模板参数不变只训练classifier这一层的参数

# Freeze parameters so we don't backprop through them
#print("model.parameters()=",type(model.parameters()))

import torch
from torch import nn
from torch import optim
import torch.nn.functional as F
from torchvision import datasets, transforms, models
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import matplotlib.pyplot as plt
import helper
from collections import OrderedDict

data_dir = 'Cat_Dog_data'

# TODO: Define transforms for the training data and testing data
train_transforms = transforms.Compose([transforms.RandomRotation(30),
                                       transforms.RandomResizedCrop(224),                                     
                                       transforms.RandomHorizontalFlip(),
                                       transforms.ToTensor(),
                                       transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))                                      
                                      ])

test_transforms = transforms.Compose([transforms.RandomRotation(30),
                                       transforms.RandomResizedCrop(224),                                       
                                       transforms.RandomHorizontalFlip(),
                                       transforms.ToTensor(),
                                       transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))                                      
                                      ])

# Pass transforms in here, then run the next cell to see how the transforms look
train_data = datasets.ImageFolder(data_dir + '/train', transform=train_transforms)
test_data = datasets.ImageFolder(data_dir + '/test', transform=test_transforms)

trainloader = torch.utils.data.DataLoader(train_data, batch_size=64, shuffle=True)
testloader = torch.utils.data.DataLoader(test_data, batch_size=64)

epoches =1
every=70
steps=0

tr_losses = []
te_losses = []
test_accu=0
train_loss=0
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") #选什么设备
model = models.densenet121(pretrained=True)
for param in model.parameters():
    param.requires_grad = False

from collections import OrderedDict
classifier = nn.Sequential(OrderedDict([
                          ('fc1', nn.Linear(1024, 500)),
                          ('relu', nn.ReLU()),
                          ('fc2', nn.Linear(500, 2)),
                          ('output', nn.LogSoftmax(dim=1))
                          ]))
    
model.classifier = classifier
criterion = nn.NLLLoss()
opti = optim.Adam(model.classifier.parameters(),lr=0.001)
model.to(device)

for e in range(epoches):

    print("trainloader.len=",len(trainloader))
    for inputs,labels in trainloader:

        opti.zero_grad()
        inputs,labels = inputs.to(device),labels.to(device)
        steps+=1
        #print("inputs.shape=",inputs.shape)
        #imgs = imgs.view(inputs.shape[0],-1)
        ps = model.forward(inputs)

       
        loss = criterion(ps,labels)
        train_loss+=loss
        tr_losses.append(loss)
        loss.backward()
        opti.step()
        
        if steps%every==0:
            test_loss=0
            test_accu=0
            model.eval()
            with torch.no_grad():
                for imgs2 ,lb2 in testloader:
                    imgs2,lb2=imgs2.to(device),lb2.to(device)
                    #imgs2 = imgs2.view(imgs2.shape[0],-1)
                    ps2=model.forward(imgs2)

                   
                    loss2 = criterion(ps2,lb2)
                    te_losses.append(loss2)
                    test_loss+=loss2

                    ps2 = torch.exp(ps2)#为了计算准确率
                    t_p,t_c = ps2.topk(1,dim=1)
                    equals = t_c==lb2.view(t_c.shape)
                    test_accu = torch.mean(equals.type(torch.FloatTensor))
            model.train()
            print("accur={}".format(test_accu),"train_loss={}".format(train_loss/every),"test_loss={}".format(test_loss/len(testloader)))               
            train_loss=0
print("######################################观察训练与校验损失函数走向便于确定过拟合位置")
plt.plot(tr_losses, label='Training loss')
plt.plot(te_losses, label='Validation loss')
plt.legend(frameon=False)     

https://classroom.udacity.com/nanodegrees/nd009-cn-advanced/parts/5f4d630c-d15a-412c-aaeb-b57ad61cd03c/modules/3aa9e812-62cd-4ae3-8fc4-593538f08455/lessons/9b014a97-2267-4f1b-af97-284b7dac2a58/concepts/572cd59e-540f-43d9-906f-33d22a4452a6