qq_42374603的博客

方法一在模型、损失函数和数据中使用cudamodel1 = model()if torch.cuda.is_available(): model1 = model1.cuda()bloss = nn.CrossEntropyLoss()if torch.cuda.is_available(): bloss = bloss.cuda()for data in train_dataloader: imgs, targets = data if to

模型训练套路——以cifar10为例1.加载数据集dataset_transform_compose = transforms.Compose([transforms.ToTensor()])train_set = torchvision.datasets.CIFAR10(root="./dataset", train=True, transform=dataset_transform_compose, download=True)test_set = torchvision.datasets.C

1. 方法一(保存模型及参数）vgg16_false = torchvision.models.vgg16(pretrained=False)##加载初始模型，模型参数还没有训练torch.save(vgg16_false, "method1.pth")torch.load("method1.pth") 2. 方法二（仅保存模型参数）torch.save(vgg16_false.state_dict(), "method2.pth")torch.load("method2.pth")注：当

1. 加载已有模型——以VGG为例vgg16_true = torchvision.models.vgg16(pretrained=True)##加载参数已经训练好的模型vgg16_false = torchvision.models.vgg16(pretrained=False)##加载初始模型，模型参数还没有训练2 . 修改模型2.1 在模型的基础上增加网络结构vgg16_false.add_module("linear", nn.Linear(1000, 10))print(vgg16_

优化器的使用1.创建优化器2.循环内梯度清零3.计算梯度4.根据梯度优化loss例子：dataset_transform_compose = transforms.Compose([transforms.ToTensor()])train_set = torchvision.datasets.CIFAR10(root="./dataset", train=True, transform=dataset_transform_compose, download=True)test_set = t

1.常见损失函数1.1 L1loss例子：x = torch.Tensor([1, 2, 3])y = torch.Tensor([1, 0, 6])L1loss = nn.L1Loss()loss1 = L1loss(x, y)print(loss1)输出：tensor(1.6667)1.2 MSELoss例子：x = torch.Tensor([1, 2, 3])y = torch.Tensor([1, 0, 6])mseloss = nn.MSELoss()los

1.卷积层（以Conv2d为例）注：输入的参数包括batchsize，通道数，图像宽度和高度，对于一般的图像通常只有宽度和高度两个参数，所以可以使用reshape函数改变尺寸。例1：以张量为例input = torch.Tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9] ])print(input.shape)input = torch.

写在最前：多看pytorch官方文档1.官方文档中给出了数据的参数及使用方法2.实战dataset_transform_compose = transforms.Compose([transforms.ToTensor()])#下载数据集train_set = torchvision.datasets.CIFAR10(root="./dataset", train=True, transform=dataset_transform_compose, download=True)test_set

1.整体框架例子：（以transforms.ToTensor()为例）from torchvision import transformsimage_path = "IMG_1098.JPG"image = Image.open(image_path)trans_to_tensor = transforms.ToTensor()image_tensor = trans_to_tensor(image)print(image_tensor)2.常见的transform（1）Normal