I. Introduction
The previous article briefly introduced the data types of pytorch. This article will mainly introduce operations such as tensor indexing and slicing.
2. Data operation
1. Index
a = torch.randn(64,3,28,28) #以MNIST的数据维度为例
[In] a[0].shape #首张输入图片的维度信息
[Out] torch.Size([3,28,28])
[In] a[0,0].shape #首张图片的第一个通道的维度信息
[Out] torch.Size([28,28])
[In] a[0,0,27,27].shape #像素值
[Out] tensor(-1.5442)
[In] a.index_select(2,torch.arange(8)).shape #在维度2的位置，数值为0~7
[Out] torch.Size([64, 3, 8, 28])b = torch.randn(2,3)
[Out]tensor([[-0.9997, -0.1196, -1.5024],
[-0.1027, -0.9639, 0.9655]])
[In] mask = b.ge(0.5) #在数组中比0.5大的数位True，反之为False
[Out] tensor([[False, False, False],
[False, False, True]])
[In] torch.masked_select(b,mask) #获取b中比0.5大的元素
[Out] tensor([0.9655])
2. Slicing
a = torch.randn(64,3,28,28)
[In] a[0:2].shape #在维度0的位置取0~1，手写体数字而言就是选取前两张图片的信息
[Out] torch.Size([2, 3, 28, 28])
[In] a[0:2,:1,:,:].shape #维度0选取0~1，维度1选取0，维度2选取0~27，维度3选取0~27
[Out] torch.Size([2, 1, 28, 28])
[In] a[:2,1:,:,:].shape #维度0选取0~1，维度1选取1，维度2选取0~27，维度3选取0~27
[Out] torch.Size([2, 2, 28 ,28])
[In] a[:2,-1:,:,:].shape #维度0选取0~1，维度1选取2，维度2选取0~27，维度3选取0~27
[Out] torch.Size([2, 1, 28, 28])
[In] a[:,:,:28:2,::2].shape #维度0全选，维度1全选，维度2在0~27步长位2选取，维度3相同
[Out] torch.Size([64, 3, 14, 14])
[In] a[...].shape #全选 a = a[...]
[Out] torch.Size([64, 3, 28, 28])
[In] a[0,...].shape #维度0取0，其他维度全选
[Out] torch.Size([3, 28, 28])
[In] a[:,1,...].shape #维度1取1，其他维度全选
[Out] torch.Size([64, 28, 28])
3. Summary
Indexing and slicing of data is simpler, the same as arrays. But when these two operations appear in the project, I will suddenly forget or confuse them, so I will record them here, and I still need a lot of exercise. Over a mountain is another mountain, see you on the next mountain.