拼接

cat

torch.cat([a,b], dim=n)：将tensor按指定的维度拼接。 合并的tensor维度要一样，除了合并以外的其他维度数据量也要一样。

In [1]: import torch

In [2]: a = torch.rand(4,32,8)

In [3]: b = torch.rand(5,32,8)

In [4]: torch.cat([a,b],dim=0).shape  # 按dim=0 合并
Out[4]: torch.Size([9, 32, 8])

stack

torch.stack([a,b],dim=n)：创建一个新的维度。
两个tensor的维度必须一摸一样

In [5]: a1 = torch.rand(4,3,16,32)

In [6]: a2 = torch.rand(4,3,16,32)

In [7]: torch.stack([a1,a2],dim=2).shape
Out[7]: torch.Size([4, 3, 2, 16, 32])

分割

split

a.split([len1,len2],dim=n)：将tensor在n维上按len1，len2 进行拆分。

In [8]: a = torch.rand(3,32,8)

In [9]: aa,bb = a.split([2,1],dim=0)

In [10]: aa.shape,bb.shape
Out[10]: (torch.Size([2, 32, 8]), torch.Size([1, 32, 8]))

In [11]: aa,bb,cc = a.split(1,dim=0)

In [12]: aa.shape,bb.shape,cc.shape
Out[12]: (torch.Size([1, 32, 8]), torch.Size([1, 32, 8]), torch.Size([1, 32, 8]))

chunk

a.chunl(n,dim=n)：在n维上将tensor分成n个

In [13]: aa,bb = a.chunk(2,dim=0)

In [14]: aa.shape
Out[14]: torch.Size([2, 32, 8])

In [15]: bb.shape
Out[15]: torch.Size([1, 32, 8])

数学计算

基本加减乘除

+ = (torch.add(a,b));
- = (torch.sub(a,b));
* = (torch.mul(a,b));
/ = (torch.div(a,b));
** = (a.pow(2))

自然指数和自然对数

In [21]: a = torch.exp(torch.ones(2,2))

In [22]: a
Out[22]:
tensor([[2.7183, 2.7183],
        [2.7183, 2.7183]])

In [23]: torch.log(a)
Out[23]:
tensor([[1., 1.],
        [1., 1.]])

In [24]: torch.log2(a)
Out[24]:
tensor([[1.4427, 1.4427],
        [1.4427, 1.4427]])

矩阵的乘法

@ = (torch.matmul(a,b)), 大于2维的tensor做矩阵的乘法，取后面两维计算，前面不变

In [16]: a = torch.rand(4,3,28,64)

In [17]: b = torch.rand(4,3,64,32)

In [18]: torch.matmul(a,b).shape
Out[18]: torch.Size([4, 3, 28, 32])

近似

a.floor()：向下近似；
a.ceil()：向上近似；
a.round()：四舍五入；
a.trunc()：取整数部分；
a.frac()：取小数部分.

clamp

a.clamp(n)：a中不到n的值变成n；
a.clamp(n,m)：a中超过m的变成m.

In [25]: a = torch.rand(2,3)*10

In [26]: a
Out[26]:
tensor([[7.2193, 0.8880, 8.3444],
        [9.9666, 6.6064, 8.0220]])

In [27]: a.max()
Out[27]: tensor(9.9666)

In [28]: a.median()
Out[28]: tensor(7.2193)

In [29]: a.clamp(8)
Out[29]:
tensor([[8.0000, 8.0000, 8.3444],
        [9.9666, 8.0000, 8.0220]])

In [30]: a.clamp(1,8)
Out[30]:
tensor([[7.2193, 1.0000, 8.0000],
        [8.0000, 6.6064, 8.0000]])

统计

norm（范数）

1-范数

$${\Vert X\Vert }_1=\sum _{i=1}^n\left|a_i\right|$$
a.norm(1,dim=n)：在第n维求1-范数

2-范数(F-范数)

$${\Vert X\Vert }_2=\sqrt{\sum _{i=1}^n{a}_i^2}$$
a.norm(2,dim=n)：在第n维求2-范数

P-范数

$${\Vert X\Vert }_2=(\sum _{i=1}^n{a}_i^p{)}^{1/p}$$
a.norm(3,dim=n)：在第n维求P-范数

mean, sum, min, max, prod

In [2]: a=torch.arange(8).view(2,4).float()

In [3]: a
Out[3]:
tensor([[0., 1., 2., 3.],
        [4., 5., 6., 7.]])

In [4]: a.min(),a.max(),a.mean(),a.prod()
Out[4]: (tensor(0.), tensor(7.), tensor(3.5000), tensor(0.))

a.prod：求累乘

argmin,argmax

如果没有指定维度，那么会将tensor先打平成维度为1的tensor然后返回索引。

In [5]: a.argmin(),a.argmax()
Out[5]: (tensor(0), tensor(7))

In [6]: a=torch.randn(4,10)

In [7]: a
Out[7]:
tensor([[-1.3142,  1.3024, -0.6175, -1.9788, -0.6462, -1.7694,  0.9113,  0.2646,
         -1.4072, -0.3260],
        [ 1.5882, -0.1465,  0.1068, -0.2997, -1.5346,  0.7147,  1.0094,  0.8033,
          0.7738, -0.7427],
        [-1.7755,  0.5528, -0.6305, -0.9959,  1.0716, -1.2245, -0.5265, -0.4162,
          0.8127,  1.2548],
        [ 0.3259,  0.4556,  0.4757, -0.7854,  0.6494, -1.2899, -0.7239,  0.7183,
         -0.3027,  0.2722]])

In [8]: a.argmax()
Out[8]: tensor(10)

In [9]: a.argmax(dim=1)
Out[9]: tensor([1, 0, 9, 7])

In [10]: a.argmax(dim=1,keepdim=True)
Out[10]:
tensor([[1],
        [0],
        [9],
        [7]])

Top-k or k-th

In [11]: a.topk(3,dim=1)  # 返回最大的三个
Out[11]:
torch.return_types.topk(
values=tensor([[1.3024, 0.9113, 0.2646],
        [1.5882, 1.0094, 0.8033],
        [1.2548, 1.0716, 0.8127],
        [0.7183, 0.6494, 0.4757]]),
indices=tensor([[1, 6, 7],
        [0, 6, 7],
        [9, 4, 8],
        [7, 4, 2]]))

In [12]: a.topk(3,dim=1,largest=False)  # 返回最小的三个
Out[12]:
torch.return_types.topk(
values=tensor([[-1.9788, -1.7694, -1.4072],
        [-1.5346, -0.7427, -0.2997],
        [-1.7755, -1.2245, -0.9959],
        [-1.2899, -0.7854, -0.7239]]),
indices=tensor([[3, 5, 8],
        [4, 9, 3],
        [0, 5, 3],
        [5, 3, 6]]))

比较

In [13]: a>0
Out[13]:
tensor([[False,  True, False, False, False, False,  True,  True, False, False],
        [ True, False,  True, False, False,  True,  True,  True,  True, False],
        [False,  True, False, False,  True, False, False, False,  True,  True],
        [ True,  True,  True, False,  True, False, False,  True, False,  True]])

In [14]: torch.gt(a,0)
Out[14]:
tensor([[False,  True, False, False, False, False,  True,  True, False, False],
        [ True, False,  True, False, False,  True,  True,  True,  True, False],
        [False,  True, False, False,  True, False, False, False,  True,  True],
        [ True,  True,  True, False,  True, False, False,  True, False,  True]])

In [15]: a!=0
Out[15]:
tensor([[True, True, True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True, True, True]])

torch.eq(a,b) ：比较两个tensor是否一样。