1 模型的保存和加载

1.1 保存与加载整个模型

保存网络的所有模块，代码量少。

但是这种方法缺点是保存模型的时候，序列化的数据被绑定到了特定的类和确切的目录。 这是因为pickle不保存模型类本身，而是保存这个类的路径， 并且在加载的时候会使用。因此， 当在其他项目里使用或者重构的时候，加载模型的时候会出错。

保存模型

import torch

torch.save(net,PATH)

加载模型

model=torch.load(PATH)

1.2 保存与加载模型中的参数

保存网络中的参数，速度快，占用空间少。

保存模型参数

torch.save(net.state_dict(),PATH)

加载模型参数

#定义模型结构
model=Model().cuda()

model.load_state_dict(torch.load(PATH))

1.3 保存与加载自定义模型

保存自定义模型参数

torch.save(
    {
        'epoch': epochID + 1,
        'state_dict': model.state_dict(),
        'best_loss': lossMIN,
        'optimizer': optimizer.state_dict(),
        'alpha': loss.alpha,
        'gamma': loss.gamma
    },
    checkpoint_path + '/m-' + launchTimestamp + '-' + str("%.4f" % lossMIN) + '.pth.tar')

比如我们在上述代码中保存了epochID，模型的state_dict，min_loss，optimizer的state_dict等。

加载自定义模型

def load_checkpoint(model, checkpoint_PATH, optimizer):
    if checkpoint != None:
        model_CKPT = torch.load(checkpoint_PATH)
        model.load_state_dict(model_CKPT['state_dict'])
        print('loading checkpoint!')
        optimizer.load_state_dict(model_CKPT['optimizer'])
    return model, optimizer

如果对网络进行了增删查改，那么需要过滤一些旧的参数，那么加载代码修改为

def load_checkpoint(model, checkpoint, optimizer, loadOptimizer):
    if checkpoint != 'No':
        print("loading checkpoint...")
        model_dict = model.state_dict()
        modelCheckpoint = torch.load(checkpoint)
        pretrained_dict = modelCheckpoint['state_dict']
        # 过滤操作
        new_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict.keys()}
        model_dict.update(new_dict)
        # 打印出来，更新了多少的参数
        print('Total : {}, update: {}'.format(len(pretrained_dict), len(new_dict)))
        model.load_state_dict(model_dict)
        print("loaded finished!")
        # 如果不需要更新优化器那么设置为false
        if loadOptimizer == True:
            optimizer.load_state_dict(modelCheckpoint['optimizer'])
            print('loaded! optimizer')
        else:
            print('not loaded optimizer')
    else:
        print('No checkpoint is included')
    return model, optimizer

1.4 在加载的模型上继续训练

在训练模型的时候可能会因为一些问题导致程序中断，或者常常需要观察训练情况的变化来更改学习率等参数，这时候就需要加载中断前保存的模型，并在此基础上继续训练

#-*- coding:utf-8 -*-

'''本文件用于举例说明pytorch保存和加载文件的方法'''

import torch as torch
import torchvision as tv
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torchvision.transforms as transforms
from torchvision.transforms import ToPILImage
import torch.backends.cudnn as cudnn
import datetime
import argparse

# 参数声明
batch_size = 32
epochs = 10
WORKERS = 0   # dataloder线程数
test_flag = True  #测试标志，True时加载保存好的模型进行测试 
ROOT = '/home/pxt/pytorch/cifar'  # MNIST数据集保存路径
log_dir = '/home/pxt/pytorch/logs/cifar_model.pth'  # 模型保存路径

# 加载MNIST数据集
transform = tv.transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])

train_data = tv.datasets.CIFAR10(root=ROOT, train=True, download=True, transform=transform)
test_data = tv.datasets.CIFAR10(root=ROOT, train=False, download=False, transform=transform)

train_load = torch.utils.data.DataLoader(train_data, batch_size=batch_size, shuffle=True, num_workers=WORKERS)
test_load = torch.utils.data.DataLoader(test_data, batch_size=batch_size, shuffle=False, num_workers=WORKERS)


# 构造模型
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 64, 3, padding=1)
        self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
        self.conv3 = nn.Conv2d(128, 256, 3, padding=1)
        self.conv4 = nn.Conv2d(256, 256, 3, padding=1)
        self.pool = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(256 * 8 * 8, 1024)
        self.fc2 = nn.Linear(1024, 256)
        self.fc3 = nn.Linear(256, 10)


    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = self.pool(F.relu(self.conv2(x)))
        x = F.relu(self.conv3(x))
        x = self.pool(F.relu(self.conv4(x)))
        x = x.view(-1, x.size()[1] * x.size()[2] * x.size()[3])
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


model = Net().cuda()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01)


# 模型训练
def train(model, train_loader, epoch):
    model.train()
    train_loss = 0
    for i, data in enumerate(train_loader, 0):
        x, y = data
        x = x.cuda()
        y = y.cuda()
        optimizer.zero_grad()
        y_hat = model(x)
        loss = criterion(y_hat, y)
        loss.backward()
        optimizer.step()
        train_loss += loss
    loss_mean = train_loss / (i+1)
    print('Train Epoch: {}\t Loss: {:.6f}'.format(epoch, loss_mean.item()))

# 模型测试
def test(model, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for i, data in enumerate(test_loader, 0):
            x, y = data
            x = x.cuda()
            y = y.cuda()
            optimizer.zero_grad()
            y_hat = model(x)
            test_loss += criterion(y_hat, y).item()
            pred = y_hat.max(1, keepdim=True)[1]
            correct += pred.eq(y.view_as(pred)).sum().item()
        test_loss /= (i+1)
        print('Test set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
            test_loss, correct, len(test_data), 100. * correct / len(test_data)))


def main():

    # 如果test_flag=True,则加载已保存的模型
    if test_flag:
        # 加载保存的模型直接进行测试机验证，不进行此模块以后的步骤
        checkpoint = torch.load(log_dir)
        model.load_state_dict(checkpoint['model'])
        optimizer.load_state_dict(checkpoint['optimizer'])
        epochs = checkpoint['epoch']
        test(model, test_load)
        return

    for epoch in range(0, epochs):
        train(model, train_load, epoch)
        test(model, test_load)
        # 保存模型
        state = {'model':model.state_dict(), 'optimizer':optimizer.state_dict(), 'epoch':epoch}
        torch.save(state, log_dir)

if __name__ == '__main__':
    main()

上述代码文件是比较常规的训练模型与保存模型文件的代码，我们可以通过保存的模型参数以及修改上述代码的main函数使其具有加载离线模型文件并可继续训练的功能，修改后的代码文件如下

#-*- coding:utf-8 -*-

'''本文件用于举例说明pytorch保存和加载文件的方法'''

import torch as torch
import torchvision as tv
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torchvision.transforms as transforms
from torchvision.transforms import ToPILImage
import torch.backends.cudnn as cudnn
import datetime
import argparse

# 参数声明
batch_size = 32
epochs = 10
WORKERS = 0   # dataloder线程数
test_flag = True  #测试标志，True时加载保存好的模型进行测试 
ROOT = '/home/pxt/pytorch/cifar'  # MNIST数据集保存路径
log_dir = '/home/pxt/pytorch/logs/cifar_model.pth'  # 模型保存路径

# 加载MNIST数据集
transform = tv.transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])

train_data = tv.datasets.CIFAR10(root=ROOT, train=True, download=True, transform=transform)
test_data = tv.datasets.CIFAR10(root=ROOT, train=False, download=False, transform=transform)

train_load = torch.utils.data.DataLoader(train_data, batch_size=batch_size, shuffle=True, num_workers=WORKERS)
test_load = torch.utils.data.DataLoader(test_data, batch_size=batch_size, shuffle=False, num_workers=WORKERS)


# 构造模型
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 64, 3, padding=1)
        self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
        self.conv3 = nn.Conv2d(128, 256, 3, padding=1)
        self.conv4 = nn.Conv2d(256, 256, 3, padding=1)
        self.pool = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(256 * 8 * 8, 1024)
        self.fc2 = nn.Linear(1024, 256)
        self.fc3 = nn.Linear(256, 10)


    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = self.pool(F.relu(self.conv2(x)))
        x = F.relu(self.conv3(x))
        x = self.pool(F.relu(self.conv4(x)))
        x = x.view(-1, x.size()[1] * x.size()[2] * x.size()[3])
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


model = Net().cuda()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01)


# 模型训练
def train(model, train_loader, epoch):
    model.train()
    train_loss = 0
    for i, data in enumerate(train_loader, 0):
        x, y = data
        x = x.cuda()
        y = y.cuda()
        optimizer.zero_grad()
        y_hat = model(x)
        loss = criterion(y_hat, y)
        loss.backward()
        optimizer.step()
        train_loss += loss
    loss_mean = train_loss / (i+1)
    print('Train Epoch: {}\t Loss: {:.6f}'.format(epoch, loss_mean.item()))

# 模型测试
def test(model, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for i, data in enumerate(test_loader, 0):
            x, y = data
            x = x.cuda()
            y = y.cuda()
            optimizer.zero_grad()
            y_hat = model(x)
            test_loss += criterion(y_hat, y).item()
            pred = y_hat.max(1, keepdim=True)[1]
            correct += pred.eq(y.view_as(pred)).sum().item()
        test_loss /= (i+1)
        print('Test set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
            test_loss, correct, len(test_data), 100. * correct / len(test_data)))


def main():

    # 如果test_flag=True,则加载已保存的模型
    if test_flag:
        # 加载保存的模型直接进行测试机验证，不进行此模块以后的步骤
        checkpoint = torch.load(log_dir)
        model.load_state_dict(checkpoint['model'])
        optimizer.load_state_dict(checkpoint['optimizer'])
        start_epoch = checkpoint['epoch']
        test(model, test_load)
        return

    # 如果有保存的模型，则加载模型，并在其基础上继续训练
    if os.path.exists(log_dir):
        checkpoint = torch.load(log_dir)
        model.load_state_dict(checkpoint['model'])
        optimizer.load_state_dict(checkpoint['optimizer'])
        start_epoch = checkpoint['epoch']
        print('加载 epoch {} 成功！'.format(start_epoch))
    else:
        start_epoch = 0
        print('无保存模型，将从头开始训练！')

    for epoch in range(start_epoch+1, epochs):
        train(model, train_load, epoch)
        test(model, test_load)
        # 保存模型
        state = {'model':model.state_dict(), 'optimizer':optimizer.state_dict(), 'epoch':epoch}
        torch.save(state, log_dir)

if __name__ == '__main__':
    main()

参考链接