网站建设服务谁便宜,大气的个人网站,广告项目网站开发,北京购物网站建设公司1. 步骤及代码
迁移学习一般都会使用两个步骤进行训练#xff1a;
固定预训练模型的特征提取部分#xff0c;只对最后一层进行训练#xff0c;使其快速收敛#xff1b;使用较小的学习率#xff0c;对全部模型进行训练#xff0c;并对每层的权重进行细微的调节。
impor…1. 步骤及代码
迁移学习一般都会使用两个步骤进行训练
固定预训练模型的特征提取部分只对最后一层进行训练使其快速收敛使用较小的学习率对全部模型进行训练并对每层的权重进行细微的调节。
import os
import torch
import torchvision
from torchvision.datasets import ImageFolder
from torch.utils.data import DataLoader
from torchvision import transforms as T
import numpy as np# 设置均值、方差
mean [0.485, 0.456, 0.406]
std [0.229, 0.224, 0.225]# 还原减均值除以方差之前的数据用于可视化
def reduction_img_show(tensor, mean, std) - None:to_img T.ToPILImage()reduced_img to_img(tensor * torch.tensor(std).view(3, 1, 1) torch.tensor(mean).view(3, 1, 1))reduced_img.show()def getResNet(*, class_names: str, loadfile: str None):if loadfile is not None:model torchvision.models.resnet18()model.load_state_dict(torch.load(resnet18-f37072fd.pth)) # 加载权重else:model torchvision.models.resnet18(weightstorchvision.models.ResNet18_Weights.IMAGENET1K_V1) # 模型自动下载到C:\Users\GaryLau\.cache\torch\hub\checkpoints# 将所有的参数层冻结设置模型除最后一层以外都不可以进行训练使模型只针对最后一层进行微调for param in model.parameters():param.requires_grad False# 输出全连接层信息print(model.fc)x model.fc.in_features # 获取全连接层输入维度model.fc torch.nn.Linear(in_featuresx, out_featureslen(class_names)) # 创建新的全连接层print(model.fc) # 输出新的全连接层return model# 定义训练函数
def train(model, device, train_loader, criterion, optimizer, epoch):model.train()all_loss []for batch_idx, (data, target) in enumerate(train_loader):data, target data.to(device), target.to(device)optimizer.zero_grad()y_pred model(data)loss criterion(y_pred, target)loss.backward()all_loss.append(loss.item())optimizer.step()if batch_idx % 10 0:print(Train Epoch: {} [{}/{}]\tLoss: {:.6f}.format(epoch, batch_idx * len(data), len(train_loader.dataset),np.mean(all_loss)))def val(model, device, val_loader, criterion):model.eval()test_loss []correct []with torch.no_grad():for data, target in val_loader:data, target data.to(device), target.to(device)y_pred model(data)test_loss.append(criterion(y_pred, target).item())pred y_pred.argmax(dim1, keepdimTrue)correct.append(pred.eq(target.view_as(pred)).sum().item()/pred.size(0))print(--Test: Average loss:{:.4f}, Accuracy:({:.0f}%)\n.format(np.mean(test_loss), 100 * sum(correct) / len(correct)))# 训练验证时的预处理
transform {train: T.Compose([T.RandomResizedCrop(224),T.RandomHorizontalFlip(),T.ToTensor(),T.Normalize(meanmean, stdstd)]),val: T.Compose([T.Resize((224,224)),T.ToTensor(),T.Normalize(meanmean, stdstd)])}# 加载训练、验证数据
dataset_train ImageFolder(r./train, transformtransform[train])
dataset_val ImageFolder(r./test, transformtransform[val])# 类别标签
class_names dataset_train.classes
print(dataset_train.class_to_idx)
print(dataset_val.class_to_idx)# 显示一张训练、验证图
# reduction_img_show(dataset_train[0][0], mean, std)
# reduction_img_show(dataset_val[0][0], mean, std)# 使用DataLoader遍历数据
dataloader_train DataLoader(dataset_train, batch_size16, shuffleTrue, samplerNone, num_workers0,pin_memoryFalse, drop_lastFalse)
dataloader_val DataLoader(dataset_val, batch_size16, shuffleFalse, samplerNone, num_workers0,pin_memoryFalse, drop_lastFalse)# 使用方式一使用next不断获取一个batch的数据
dataiter_train iter(dataloader_train)
imgs, labels next(dataiter_train)
print(imgs.size())
# reduction_img_show(imgs[0], mean, std)
# reduction_img_show(imgs[1], mean, std)
multi_imgs torchvision.utils.make_grid(imgs, nrow10) # 拼接一个batch的图像用于展示
# reduction_img_show(multi_imgs, mean, std)# 获取ResNet模型并加载预训练模型权重将最后一层(输出层)去掉换成一个新的全连接层新全连接层输出的节点数是新数据的类别数
device torch.device(cuda:0 if torch.cuda.is_available() else cpu)
print(device)# 构建模型
model getResNet(class_namesclass_names, loadfileresnet18-f37072fd.pth)
model.to(device)# 构建损失函数
criterion torch.nn.CrossEntropyLoss()
# 指定新加的全连接层为要更新的参数
optimizer torch.optim.Adam(model.fc.parameters(), lr0.001) # 只需要更新最后一层fc的参数if __name__ __main__:### 步骤一微调最后一层first_model resnet18-f37072fd_finetune_fcLayer.pthfor epoch in range(1, 6):train(model, device, dataloader_train, criterion, optimizer, epoch)val(model, device, dataloader_val, criterion)# 仅保存了最后新添加的全连接层的参数#torch.save(model.fc.state_dict(), first_model)torch.save(model.state_dict(), first_model)### 步骤二小学习率微调所有层second_model resnet18-f37072fd_finetune_allLayer.pthoptimizer2 torch.optim.SGD(model.parameters(), lr0.001, momentum0.9)exp_lr_scheduler torch.optim.lr_scheduler.StepLR(optimizer2, step_size3, gamma0.9)# 将所有的参数层设为可训练的for param in model.parameters():param.requires_grad Trueif os.path.exists(second_model):model.load_state_dict(torch.load(second_model)) # 加载本地模型else:model.load_state_dict(torch.load(first_model)) # 加载步骤一训练得到的本地模型print(Finetune all layers with small learning rate......)for epoch in range(1, 101):train(model, device, dataloader_train, criterion, optimizer2, epoch)if optimizer2.state_dict()[param_groups][0][lr] 0.00001:exp_lr_scheduler.step()print(flearning rate: {optimizer2.state_dict()[param_groups][0][lr]})val(model, device, dataloader_val, criterion)# 保存整个模型torch.save(model.state_dict(), second_model)print(Done.)2. 完整资源
https://download.csdn.net/download/liugan528/89833913
