Tensor的操作1(索引、切片、维度变换)
目录
- 索引
-
- select by index
- 选择性的获取元素
- 对指定维度采样指定的元素
- 省略性写法
- 带mask的选法
- 先将tensor打平再选择
- Tensor的维度变换
-
- view
- squeeze and unsqueeze
-
- unsqueeze
- squeeze
- repeat and expand
- transpose and permute
-
- transpose
- permute
索引
select by index
In [1]: import torch
In [2]: a = torch.rand(4,3,28,28) # 随机生成一个4*3*28*28的tensor
In [3]: a[0].shape # 第0维的第1个元素的size
Out[3]: torch.Size([3, 28, 28])
In [4]: a[0,0].shape
Out[4]: torch.Size([28, 28])
In [5]: a[0,0,2,4]
Out[5]: tensor(0.4495)
选择性的获取元素
In [6]: a.shape
Out[6]: torch.Size([4, 3, 28, 28])
In [7]: a[:2].shape # 从开头取到2（不包括2）
Out[7]: torch.Size([2, 3, 28, 28])
In [8]: a[:2,:1].shape
Out[8]: torch.Size([2, 1, 28, 28])
In [9]: a[:2,1:].shape
Out[9]: torch.Size([2, 2, 28, 28])
In [10]: a[:2,-1:].shape # 最后一个元素取到最末尾
Out[10]: torch.Size([2, 1, 28, 28])
In [11]: a[:,:,0:28:2,0:28:2].shape # 取元素的step为2
Out[11]: torch.Size([4, 3, 14, 14])
对指定维度采样指定的元素
In [12]: a.index_select(0, torch.tensor([0,2])).shape # 对第0维采样0，2
Out[12]: torch.Size([2, 3, 28, 28])
In [13]: a.index_select(1, torch.tensor([0,2])).shape
Out[13]: torch.Size([4, 2, 28, 28])
In [14]: a.index_select(2,torch.arange(28)).shape
Out[14]: torch.Size([4, 3, 28, 28])
省略性写法
In [22]: a[...].shape
Out[22]: torch.Size([4, 3, 28, 28])
In [23]: a[0,...].shape
Out[23]: torch.Size([3, 28, 28])
In [24]: a[:,1,...].shape
Out[24]: torch.Size([4, 28, 28])
带mask的选法
In [27]: x = torch.randn(3,4)
In [28]: x
Out[28]:
tensor([[ 0.1600, 1.1167, -0.5031, -2.0422],
[ 0.0912, -0.7644, 0.8397, -0.3581],
[ 1.1684, -0.9302, -1.4750, 0.0792]])
In [29]: mask = x.ge(0.5)
In [30]: mask
Out[30]:
tensor([[False, True, False, False],
[False, False, True, False],
[ True, False, False, False]])
In [31]: torch.masked_select(x,mask)
Out[31]: tensor([1.1167, 0.8397, 1.1684])
先将tensor打平再选择
In [32]: src = torch.tensor([4,3,5],[6,7,8])
In [35]: torch.take(src,torch.tensor([0,2,5]))
Out[35]: tensor([4, 5, 8])
Tensor的维度变换
view
In [1]: import torch
In [2]: a=torch.rand(4,1,28,28)
In [3]: a.shape
Out[3]: torch.Size([4, 1, 28, 28])
In [4]: a.view(4,28*28)
Out[4]:
tensor([[0.3681, 0.6182, 0.5406, ..., 0.3441, 0.7353, 0.7404],
[0.0606, 0.1527, 0.5305, ..., 0.3635, 0.2176, 0.1675],
[0.2906, 0.9702, 0.5708, ..., 0.5743, 0.3238, 0.8871],
[0.4259, 0.3133, 0.3642, ..., 0.1067, 0.6268, 0.1117]])
In [5]: a.view(4,28*28).shape
Out[5]: torch.Size([4, 784])
In [6]: a.view(4*28,28).shape
Out[6]: torch.Size([112, 28])
注意：
1.view变换前后的总体size要相等
2.会丢失数据的维度信息，不方便恢复
# 维度信息有问题的报错
In [8]: a.view(4,743)
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-8-9ba94a9550d8> in <module>
----> 1 a.view(4,743)
RuntimeError: shape '[4, 743]' is invalid for input of size 3136
squeeze and unsqueeze
unsqueeze
在指定的位置增加一个维度
注意：Tensor从头到尾的index是从0开始，从尾到头的index是从-1开始
In [12]: a.shape
Out[12]: torch.Size([4, 1, 28, 28])
In [13]: a.unsqueeze(0).shape
Out[13]: torch.Size([1, 4, 1, 28, 28])
In [14]: a.unsqueeze(-1).shape
Out[14]: torch.Size([4, 1, 28, 28, 1])
In [15]: a.unsqueeze(3).shape
Out[15]: torch.Size([4, 1, 28, 1, 28])
In [16]: a.unsqueeze(-4).shape
Out[16]: torch.Size([4, 1, 1, 28, 28])
In [17]: a.unsqueeze(5).shape
---------------------------------------------------------------------------
IndexError Traceback (most recent call last)
<ipython-input-17-b54eab361a50> in <module>
----> 1 a.unsqueeze(5).shape
IndexError: Dimension out of range (expected to be in range of [-5, 4], but got 5)
In [18]:
如果要将两个维度不一样的tensor相加则要用到unsqueeze
In [18]: b = torch.rand(32)
In [19]: f = torch.rand(4,32,14,14)
In [22]: b= b.unsqueeze(1).unsqueeze(2).unsqueeze(0)
In [23]: b.shape
Out[23]: torch.Size([1, 32, 1, 1])
squeeze
In [24]: b.shape
Out[24]: torch.Size([1, 32, 1, 1])
In [25]: b.squeeze().shape # 把能压缩的都压缩了
Out[25]: torch.Size([32])
In [26]: b.squeeze(0).shape
Out[26]: torch.Size([32, 1, 1])
In [27]: b.squeeze(-2).shape
Out[27]: torch.Size([1, 32, 1])
repeat and expand
expand：不主动复制数据
`repeat``：直接复制数据
In [34]: a=torch.rand(4,32,14,14)
In [35]: b.shape
Out[35]: torch.Size([1, 32, 1, 1])
In [36]: b.expand(4,32,14,14).shape
Out[36]: torch.Size([4, 32, 14, 14])
repeat表示拷贝数据的次数
In [37]: b.shape
Out[37]: torch.Size([1, 32, 1, 1])
In [38]: b.repeat(4,32,1,2).shape
Out[38]: torch.Size([4, 1024, 1, 2])
transpose and permute
transpose
a.transpose(d1,d2)只能转换两个维度
In [39]: a=torch.rand(4,3,28,28)
In [40]: a.transpose(1,3).shape
Out[40]: torch.Size([4, 28, 28, 3])
permute
随意变化多个维度
In [41]: a=torch.rand(4,3,28,28)
In [42]: a.shape
Out[42]: torch.Size([4, 3, 28, 28])
In [43]: a.permute(0,3,2,1).shape
Out[43]: torch.Size([4, 28, 28, 3])