pytorch_basics

BUG

• BUG1

  Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor for argument #2 ‘weight’

是因为我的x, y两个Tensor是.cuda()的矩阵，而Linear不是.cuda()的。所以得给linear也加上.cuda()

• BUG2

  TypeError: can’t convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

看错误提示就知道了。

Tutorial

• 1. Basic autograd example 1 (Line 25 to 39)

     torch.tensor(.cuda())
     y.backward()
     x.grad

• 1. Basic autograd example 2 (Line 46 to 83)
     

     torch.randn
     torch.nn.Linear(3,2)
     (No used)linear.weight.data.sub_(0.01 * linear.weight.grad.data)
     Build loss function and optimizer.
     nn.MSELoss()
     torch.optim.SGD(linear.parameters(), lr = 0.01)
     详细不写了，写出了步骤
     Forward pass
     Compute loss
     Backward pass.
     loss.backward()
     Print out the gradients.
     1-step gradient descent.
     optimizer.step()

• 1. Loading data from numpy (Line 90 to 97)

     torch.from_numpy().cuda()
     y.cpu().numpy()

• 1. Input pipline (Line 104 to 129)

     torchvision.datasets.CIFAR10(...)
     # Data loader
     torch.utils.data.DataLoader(..)
     # Actual usage of the data loader is as below.
     for images, labels in train_loader:
     # Training code should be written here.
     pass
     

• 1. Input pipline for custom dataset (Line 136 to 156)

     build your custom dataset.
     _init__存放文件名称列表
     _getitem_读取单个用例，预处理，返回一个data和一个label
     _len_# you should change 0 to the total size of your dataset

• 1. Pretrained model (Line 163 to 176)

     resnet = torchvision.models.resnet18(pretrained=True)
     for param in resnet.parameters():
     param.requires_grad = False
     #Replace the top layer for funetuning
     resnet.fc = nn.Linear(resnet.fc.in_features, 100)

• 1. Save and load model (Line 183 to 189)

     torch.save(model,'_.ckpt') / torch.save(model.state_dict(),'_.ckpt')
     torch.load() / model.load_state_dict()