【技术分享】pytorch的FINETUNING实践（resnet18 cifar10)
- 2019 年 12 月 1 日
- 筆記
本文主要是用pytorch训练resnet18模型，对cifar10进行分类，然后将cifar10的数据进行调整，加载已训练好的模型，在原有模型上FINETUNING 对调整的数据进行分类， 可参考pytorch官网教程
resnet18模型
pytorch的resnet18模型引用：https://github.com/kuangliu/pytorch-cifar
模型详情可参考github里面的models/resnet.py， 这里不做详细的说明，readme描述准确率可达到93.02%，但我本地测试迭代200次没有达到这个数字，本地200次迭代准确率为87.40%。
导入需要的包
import os import numpy as np import torch.backends.cudnn as cudnn import torch.optim as optim import torchvision import torchvision.transforms as transforms from models import * from utils import progress_bar
设置随机种子，让结果可复现
这里尝试了比较久，在cpu上运行，只需要设置torch.manual_seed(SEED)即可稳定复现结果，但在GPU上始终不行，总存在randomness的问题，后来在友人的帮助下，查了官方的资料，终于解决了这个问题，感谢。其中tensorflow在GPU似乎做不到结果可稳定复现，如果有知道的同学，还请不吝指导~
SEED = 0 torch.manual_seed(SEED) torch.cuda.manual_seed(SEED) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False np.random.seed(SEED)
设置是运行在cpu上还是gpu上
根据是否有gpu可用选择运行的设备，注意驱动的安装，版本的兼容性，驱动也折磨了我很久。。由于我运行在docker中，下载的驱动版本不一致，导致一直检测不到gpu
device = 'cuda' if torch.cuda.is_available() else 'cpu' best_acc = 0 start_epoch = 0
数据加载及预处理
数据存放在py文件同级目录下的data文件夹下，如果数据不存在，download设置的为True，会自动从pytorch上进行下载，这里对数据进行不同的转换，增加数据多样性。
transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform_train) trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=True, num_workers=2) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform_test) testloader = torch.utils.data.DataLoader(testset, batch_size=100, shuffle=False, num_workers=2)
对数据集进行调整
原来cifar数据集包含10个类别
['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
需要实践FINETUNING，所以对数据集进行了改造，由10类改为2类，分别为动物和运输工具。马算不算交通工具呢？^.^
clz_idx = trainset.class_to_idx clz_to_idx = {'animal': 0, 'transport': 1} clz = ['animal', 'transport'] animal_name = ["bird", "cat", "deer", "dog", "frog", "horse"] animal = [clz_idx[x] for x in animal_name] trainset.targets = [0 if x in animal else 1 for x in trainset.targets] trainset.class_to_idx = clz_to_idx trainset.classes = clz testset.targets = [0 if x in animal else 1 for x in testset.targets] testset.class_to_idx = clz_to_idx testset.classes = clz
加载预训练的模型
模型存放在checkpoint目录下，模型的训练是上述的Resnet18, 注意如果是gpu训练，尤其关注一下if中代码的顺序。
- 将net装换为DataParallel，用以并行训练，因为原Resnet18在gpu上训练使用了DataParallel，所以这里也要进行封装，会包一层module
- FINETUNING：将最后一层的10类输出，改为2类输出。注意gpu中的写法，net.module.linear
net = net.to(device)修改了模型之后，要将模型推送到gpu上，这步不能提前，会出现参数不在GPU上的错误
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!' checkpoint = torch.load('./checkpoint/ckpt.pth') net = ResNet18() if device == 'cuda': net = torch.nn.DataParallel(net) net.load_state_dict(checkpoint['net']) net.module.linear = nn.Linear(net.module.linear.in_features, 2) else: net.load_state_dict(checkpoint['net']) net.linear = nn.Linear(net.linear.in_features, 2) net = net.to(device)
指定不需要调整的层数
指定前40层的参数固定，不需要再学习
for idx, (name, param) in enumerate(net.named_parameters()): if idx > 40: # count of layers is 62 param.requires_grad = False if param.requires_grad == True: print("t", idx, name)
loss函数和优化算法
criterion = nn.CrossEntropyLoss() optimizer = optim.SGD(net.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4)
训练函数 及 测试函数
参考Resnet18中的main.py， 在测试的时候，保存训练的结果，用以后续继续训练，区分文件夹保存， 同时只有在精度提高的基础上进行保存。
def train(epoch): print('nEpoch: %d' % epoch) net.train() train_loss = 0 correct = 0 total = 0 for batch_idx, (inputs, targets) in enumerate(trainloader): inputs, targets = inputs.to(device), targets.to(device) optimizer.zero_grad() outputs = net(inputs) loss = criterion(outputs, targets) loss.backward() optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() # print('%d/%d, [Loss: %.03f | Acc: %.3f%% (%d/%d)]' # % (batch_idx+1, len(trainloader), train_loss/(batch_idx+1), 100.*correct/total, correct, total)) progress_bar(batch_idx, len(trainloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % (train_loss / (batch_idx + 1), 100. * correct / total, correct, total)) best_acc = 0 def test(epoch): global best_acc net.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.to(device), targets.to(device) outputs = net(inputs) loss = criterion(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % (test_loss / (batch_idx + 1), 100. * correct / total, correct, total)) # Save checkpoint. acc = 100. * correct / total if acc > best_acc: print('Saving..') state = { 'net': net.state_dict(), 'acc': acc, 'epoch': epoch, } if not os.path.isdir('checkpoint_ft'): os.mkdir('checkpoint_ft') torch.save(state, './checkpoint_ft/ckpt.pth') best_acc = acc
开始训练
由于在已经训练好的模型的基础上训练，这里的迭代次数不用太多即可以达到较高的准确率
for epoch in range(start_epoch, start_epoch + 20): train(epoch) test(epoch)
结果展示
Epoch: 0 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.520 | Acc: 88.662% (44331/50000) [================================================================>] Step: 21ms | Tot 100/100 | Loss: 0.449 | Acc: 95.090% (9509/10000) Saving.. Epoch: 1 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.430 | Acc: 95.342% (47671/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.411 | Acc: 95.590% (9559/10000) Saving.. Epoch: 2 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.394 | Acc: 95.816% (47908/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.373 | Acc: 96.110% (9611/10000) Saving.. Epoch: 3 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.376 | Acc: 96.002% (48001/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.386 | Acc: 94.560% (9456/10000) Epoch: 4 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.368 | Acc: 96.160% (48080/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.365 | Acc: 96.350% (9635/10000) Saving.. Epoch: 5 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.362 | Acc: 96.214% (48107/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.381 | Acc: 93.430% (9343/10000) Epoch: 6 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.360 | Acc: 96.070% (48035/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.362 | Acc: 95.400% (9540/10000) Epoch: 7 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.358 | Acc: 96.062% (48031/50000) [================================================================>] Step: 21ms | Tot 100/100 | Loss: 0.400 | Acc: 90.730% (9073/10000) Epoch: 8 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.356 | Acc: 96.214% (48107/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.362 | Acc: 96.280% (9628/10000) Epoch: 9 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.353 | Acc: 96.242% (48121/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.376 | Acc: 94.590% (9459/10000) Epoch: 10 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.352 | Acc: 96.348% (48174/50000) [================================================================>] Step: 21ms | Tot 100/100 | Loss: 0.384 | Acc: 93.080% (9308/10000) Epoch: 11 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.351 | Acc: 96.236% (48118/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.356 | Acc: 95.480% (9548/10000) Epoch: 12 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.350 | Acc: 96.348% (48174/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.383 | Acc: 93.170% (9317/10000) Epoch: 13 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.348 | Acc: 96.358% (48179/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.373 | Acc: 93.330% (9333/10000) Epoch: 14 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.347 | Acc: 96.446% (48223/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.391 | Acc: 91.670% (9167/10000) Epoch: 15 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.346 | Acc: 96.324% (48162/50000) [================================================================>] Step: 21ms | Tot 100/100 | Loss: 0.347 | Acc: 95.880% (9588/10000) Epoch: 16 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.344 | Acc: 96.488% (48244/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.343 | Acc: 95.980% (9598/10000) Epoch: 17 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.344 | Acc: 96.416% (48208/50000) [================================================================>] Step: 21ms | Tot 100/100 | Loss: 0.344 | Acc: 95.890% (9589/10000) Epoch: 18 [================================================================>] Step: 54ms | Tot: 3 391/391 Loss: 0.344 | Acc: 96.370% (48185/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.354 | Acc: 95.060% (9506/10000) Epoch: 19 [================================================================>] Step: 53ms | Tot: 3 391/391 Loss: 0.344 | Acc: 96.338% (48169/50000) [================================================================>] Step: 20ms | Tot 100/100 | Loss: 0.399 | Acc: 89.760% (8976/10000)
在已有准确率为87.4%的Resnet18模型上进行FINETUNING二分类，第一次迭代准确率就能达到95.09%，收敛速度还是很快的，分类效果也不错。
最终20次迭代测试集最高为96.11%。
最后
pytorch构建模型比较简单，代码看起来也很清晰，文档支持的比较全面。
