醉后不知天在水，满船清梦压星河

0%

Pytorch常用操作

发表于 分类于
本文字数： 11k 阅读时长 ≈ 10 分钟

Pytorch常用操作记录，pytorch文档：

1
https://pytorch.org/docs/stable/

基本数据类型

data type

Data type dtype CPU tensor GPU tensor
32-bit floating point torch.float32 or torch.float torch.FloatTensor torch.cuda.FloatTensor
64-bit floating point torch.float64 or torch.double torch.DoubleTensor torch.cuda.DoubleTensor
16-bit floating point torch.float16 or torch.half torch.HalfTensor torch.cuda.HalfTensor
8-bit integer (unsigned) torch.uint8 torch.ByteTensor torch.cuda.ByteTensor
8-bit integer (signed) torch.int8 torch.CharTensor torch.cuda.CharTensor
16-bit integer (signed) torch.int16 or torch.short torch.ShortTensor torch.cuda.ShortTensor
32-bit integer (signed) torch.int32 or torch.int torch.IntTensor torch.cuda.IntTensor
64-bit integer (signed) torch.int64 or torch.long torch.LongTensor torch.cuda.LongTensor
Boolean torch.bool torch.BoolTensor torch.cuda.BoolTensor

type check

1
>>> a = torch.randn(2,3)
2
>>> a.type()
3
'torch.FloatTensor'
4
5
>>> type(a)
6
torch.Tensor
7
8
>>> isinstance(a, torch.FloatTensor)
9
True

dim check

1
>>> a = torch.rand(1,2,3)
2
>>> a.shape
3
torch.Size([1,2,3])
4
5
>>> list(a.shape)
6
[1,2,3]
7
8
>>> a.size(0)
9
1
10
11
>>> a.dim()
12
3

创建Tensor

from numpy

1
>>> a = np.array([2,3.3])
2
>>> torch.from_numpy(a)
3
tensor([2, 3])

from list

1
In [*]: torch.tensor([2,3])
2
Out[*]: tensor([2, 3])
3
4
In [*]: torch.FloatTensor([2,3.3])
5
Out[*]: tensor([2.0000, 3.3000])

uninitialize

1
In [*]: torch.Tensor(2,3)
2
Out[*]:
3
tensor([[-4.2107e-05,  4.5766e-41,  7.6038e-30],
4
        [ 3.0963e-41,  1.1988e-20,  4.5766e-41]])

设置默认数据类型

1
In [*]: torch.tensor([1.2, 3]).type()
2
Out[*]: 'torch.FloatTensor'
3
4
In [*]: torch.set_default_tensor_type(torch.DoubleTensor)
5
6
In [*]: torch.tensor([1.2, 3]).type()
7
Out[*]: 'torch.DoubleTensor'

随机初始化

1
In [*]: a = torch.rand(2,2)
2
In [*]: a
3
Out[*]: tensor([[0.0934, 0.9568],
4
       			 [0.3287, 0.5201]])
5
6
In [*]: torch.rand_like(a)
7
Out[*]: tensor([[0.1323, 0.4544],
8
       		     [0.3094, 0.7894]])
9
10
In [*]: torch.randint(1,10,[2,2])
11
Out[*]: tensor([[8, 9],
12
        		 [3, 1]])
13
14
In [*]: torch.randn(2,2) #标准正态分布
15
Out[*]: tensor([[-0.6804,  0.0677],
16
       			 [-0.8232, -0.8727]])
17
18
In [*]: torch.full([2,2],5) 
19
Out[*]: tensor([[5., 5.],
20
        		 [5., 5.]])

序列初始化

1
In [*]: torch.arange(0,4)
2
Out[*]: tensor([0, 1, 2, 3])
3
4
In [*]: torch.arange(0,4,2)
5
Out[*]: tensor([0, 2])
6
7
In [*]: torch.arange(0,10)
8
Out[*]: tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
9
10
In [*]: torch.arange(0,10,2)
11
Out[*]: tensor([0, 2, 4, 6, 8])
12
13
In [*]: torch.range(0,10)
14
/home/haha/anaconda3/envs/py36/bin/ipython:1: UserWarning: torch.range is deprecated in favor of torch.arange and will be removed in 0.5. Note that arange generates values in [start; end), not [start; end].
15
  #!/home/haha/anaconda3/envs/py36/bin/python
16
Out[*]: tensor([ 0.,  1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9., 10.])
17
18
In [*]: torch.linspace(0,10,steps=5)
19
Out[*]: tensor([ 0.0000,  2.5000,  5.0000,  7.5000, 10.0000])
20
21
In [*]: torch.logspace(0,1,steps=5)
22
Out[*]: tensor([ 1.0000,  1.7783,  3.1623,  5.6234, 10.0000])

特殊初始化

1
In [*]: torch.ones(2,2)
2
Out[*]:
3
tensor([[1., 1.],
4
        [1., 1.]])
5
6
In [*]: torch.zeros(2,2)
7
Out[*]:
8
tensor([[0., 0.],
9
        [0., 0.]])
10
11
In [*]: torch.eye(2,3)
12
Out[*]:
13
tensor([[1., 0., 0.],
14
        [0., 1., 0.]])
15
16
In [*]: a = torch.zeros(3,3)
17
18
In [*]: torch.ones_like(a)
19
Out[*]:
20
tensor([[1., 1., 1.],
21
        [1., 1., 1.],
22
        [1., 1., 1.]])
23
24
In [*]: torch.randperm(5)
25
Out[*]: tensor([0, 2, 3, 1, 4])

切片和索引

indexing

1
In [*]: a = torch.rand(4,3,28,28)
2
3
In [*]: a[*].shape
4
Out[*]: torch.Size([3, 28, 28])
5
6
In [*]: a[0,0,2].shape
7
Out[*]: torch.Size([*])
8
9
In [*]: a[0,0,2,4]
10
Out[*]: tensor(0.0452)

first/last N

1
In [*]: a[:2,1:,-1:,:].shape
2
Out[*]: torch.Size([2, 2, 1, 28])

by step

1
In [*]: a[:,:,0:28:2,0:28:3].shape
2
Out[*]: torch.Size([4, 3, 14, 10])
3
4
In [*]: a[:,:,::2,::-1].shape
5
-----------------------------------------------------------------------
6
ValueError                            Traceback (most recent call last)
7
<ipython-input-73-f808e66143f0> in <module>
8
----> 1 a[:,:,::2,::-1].shape
9
10
ValueError: negative step not yet supported
11
12
In [*]: a[:,:,::2,::3].shape
13
Out[*]: torch.Size([4, 3, 14, 10])

by specific index

1
In [*]: a.index_select(1,torch.arange(2)).shape
2
Out[*]: torch.Size([4, 2, 28, 28])
3
4
In [*]: a.index_select(3,torch.arange(7)).shape
5
Out[*]: torch.Size([4, 3, 28, 7])
6
7
In [*]: a[...,:2].shape
8
Out[*]: torch.Size([4, 3, 28, 2])
9
10
In [*]: a[1,...].shape
11
Out[*]: torch.Size([3, 28, 28])

by mask

1
In [*]: x = torch.randn(3,4)
2
3
In [*]: mask = x.ge(0.5)
4
5
In [*]: mask
6
Out[*]:
7
tensor([[0, 1, 0, 0],
8
        [1, 0, 0, 0],
9
        [0, 0, 0, 0]], dtype=torch.uint8)
10
11
In [*]: torch.masked_select(x,mask)
12
Out[*]: tensor([0.7164, 2.1624])

by flatten index

1
In [*]: a = torch.tensor([[4,3,5],[6,8,7]])
2
3
In [*]: torch.take(a,torch.tensor([0,2,-1]))
4
Out[*]: tensor([4, 5, 7])

维度变换

view/reshape

1
In [*]: a = torch.rand(4,1,28,28
2
In [*]: a.view(4,28*28).shape
3
Out[*]: torch.Size([4, 784])

squeeze/unsqueeze

squeeze

1
In [*]: a.shape
2
Out[*]: torch.Size([4, 1, 28, 28])
3
4
In [*]: a.unsqueeze(0).shape
5
Out[*]: torch.Size([1, 4, 1, 28, 28])
6
7
In [*]: a.unsqueeze(-1).shape
8
Out[*]: torch.Size([4, 1, 28, 28, 1])

unsqueeze

1
In [*]: a = torch.rand(1,32,1,1)
2
3
In [*]: a.shape
4
Out[*]: torch.Size([1, 32, 1, 1])
5
6
In [*]: a.squeeze().shape
7
Out[*]: torch.Size([*])
8
9
In [*]: a.squeeze(0).shape
10
Out[*]: torch.Size([32, 1, 1])
11
12
In [*]: a.squeeze(1).shape
13
Out[*]: torch.Size([1, 32, 1, 1])

transpose/t/permute

t

only expects a 2D tensor

1
In [*]: a = torch.rand(3,4)
2
3
In [*]: a.t().shape
4
Out[*]: torch.Size([4, 3])

transpose

1
In [*]: a = torch.rand(4,3,32,32)
2
3
In [*]: a.transpose(1,3).view(4,3*32*32).shape
4
---------------------------------------------------------------
5
RuntimeError                  Traceback (most recent call last)
6
<ipython-input-45-bdff182163c0> in <module>
7
----> 1 a.transpose(1,3).view(4,3*32*32).shape
8
9
RuntimeError: invalid argument 2: view size is not compatible with input tensor size and stride (at least one dimension spans across two contiguous subspaces). Call .contiguous() before .view(). at /pytorch/aten/src/TH/generic/THTensor.cpp:203
10
11
In [*]: a1 = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,3,32,32)
12
13
In [*]: a2 = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,32,32,3).transpose(1,3)
14
15
In [*]: a1.shape,a2.shape
16
Out[*]: (torch.Size([4, 3, 32, 32]), torch.Size([4, 3, 32, 32]))
17
18
In [*]: torch.all(torch.eq(a,a1))
19
Out[*]: tensor(0, dtype=torch.uint8)
20
21
In [*]: torch.all(torch.eq(a,a2))
22
Out[*]: tensor(1, dtype=torch.uint8)

permute

1
In [*]: a.permute(0,2,3,1).shape
2
Out[*]: torch.Size([4, 32, 32, 3])

expand/repeat

expand: broadcasting

1
In [*]: a = torch.rand(1,32,1,1)
2
3
In [*]: a.expand(4,32,14,14).shape
4
Out[*]: torch.Size([4, 32, 14, 14])

repeat: memory copied

1
In [*]: a = torch.rand(1,32,1,1)
2
3
In [*]: a.repeat(4,32,1,1).shape
4
Out[*]: torch.Size([4, 1024, 1, 1])
5
6
In [*]: a.repeat(4,1,1,1).shape
7
Out[*]: torch.Size([4, 32, 1, 1])
8
9
In [*]: a.repeat(4,1,32,32).shape
10
Out[*]: torch.Size([4, 32, 32, 32])

拼接与拆分

cat

1
In [*]: a = torch.rand(4,32,8)
2
3
In [*]: b = torch.rand(5,32,8)
4
5
In [*]: torch.cat([a,b],dim=0).shape
6
Out[*]: torch.Size([9, 32, 8])

stack

create new dim, 两个tensor维度相同

1
In [*]: a.shape
2
Out[*]: torch.Size([4, 32, 8])
3
4
In [*]: torch.stack([a,a],dim=1).shape
5
Out[*]: torch.Size([4, 2, 32, 8])
6
7
In [*]: torch.stack([a,a],dim=3).shape
8
Out[*]: torch.Size([4, 32, 8, 2])

split

by length

1
In [*]: c = torch.rand(2,32,8)
2
3
In [*]: a,b = c.split([10,22],dim=1)
4
5
In [*]: a.shape,b.shape
6
Out[*]: (torch.Size([2, 10, 8]), torch.Size([2, 22, 8]))
1
In [*]: a,b = c.split(16,dim=1)
2
3
In [*]: a.shape,b.shape
4
Out[*]: (torch.Size([2, 16, 8]), torch.Size([2, 16, 8]))
5
6
In [*]: a = c.split(2,dim=0)
7
8
In [*]: a.shape
9
---------------------------------------------------------------
10
AttributeError                Traceback (most recent call last)
11
<ipython-input-86-d74f1bcdd37c> in <module>
12
----> 1 a.shape
13
14
AttributeError: 'tuple' object has no attribute 'shape'
15
16
In [*]: len(a)
17
Out[*]: 1
18
19
In [*]: a[*].shape
20
Out[*]: torch.Size([2, 32, 8])

chunk

by num

1
In [*]: a,b = c.chunk(2,dim=0)
2
3
In [*]: a.shape,b.shape
4
Out[*]: (torch.Size([1, 32, 8]), torch.Size([1, 32, 8]))

数学运算

add/minus/multiply/divide

1
In [*]: a = torch.rand(3,4)
2
3
In [*]: b = torch.rand(4)
4
5
In [*]: a+b
6
Out[*]:
7
tensor([[0.8015, 1.1462, 1.0086, 0.3205],
8
        [1.2817, 1.0414, 1.3938, 0.8069],
9
        [0.4473, 0.4525, 1.7025, 0.9032]])
10
11
In [*]: torch.add(a,b)
12
Out[*]:
13
tensor([[0.8015, 1.1462, 1.0086, 0.3205],
14
        [1.2817, 1.0414, 1.3938, 0.8069],
15
        [0.4473, 0.4525, 1.7025, 0.9032]])

matmul

torch.mm: only for 2d tensor

1
In [*]: a
2
Out[*]:
3
tensor([[3., 3.],
4
        [3., 3.]])
5
6
In [*]: b = torch.ones(2,2)
7
8
In [*]: torch.mm(a,b)
9
Out[*]:
10
tensor([[6., 6.],
11
        [6., 6.]])
12
13
In [*]: torch.matmul(a,b)
14
Out[*]:
15
tensor([[6., 6.],
16
        [6., 6.]])
17
18
In [*]: a@b
19
Out[*]:
20
tensor([[6., 6.],
21
        [6., 6.]])

floor/ceil/round/trunc/frac

1
In [*]: a
2
Out[*]: tensor(3.1400)
3
4
In [*]: a.floor(),a.ceil(),a.trunc(),a.frac()
5
Out[*]: (tensor(3.), tensor(4.), tensor(3.), tensor(0.1400))
6
7
In [*]: torch.tensor(3.499).round()
8
Out[*]: tensor(3.)
9
10
In [*]: torch.tensor(3.5).round()
11
Out[*]: tensor(4.)

clamp

1
In [*]: a
2
Out[*]:
3
tensor([[8.0798, 8.1040, 3.2681],
4
        [2.9466, 3.6986, 7.1211]])
5
6
In [*]: a.clamp(3.5,7.5)
7
Out[*]:
8
tensor([[7.5000, 7.5000, 3.5000],
9
        [3.5000, 3.6986, 7.1211]])

统计属性

min/max/mean/sum/argmax/argmin

1
In [*]: a
2
Out[*]:
3
tensor([[0., 1., 2., 3.],
4
        [4., 5., 6., 7.]])
5
6
In [*]: a.min(),a.max(),a.mean(),a.prod(),a.sum(),a.argmax(),a.argmin()
7
Out[*]:
8
(tensor(0.),
9
 tensor(7.),
10
 tensor(3.5000),
11
 tensor(0.),
12
 tensor(28.),
13
 tensor(7),
14
 tensor(0))
15
16
In [*]: a.argmax(dim=1)
17
Out[*]: tensor([3, 3]) 
18
19
In [*]: a.argmax(dim=1, keepdim=True)
20
Out[*]:
21
tensor([[*],
22
        [*]])
23
24
In [*]: a.max(dim=1)
25
Out[*]:
26
torch.return_types.max(
27
values=tensor([3., 7.]),
28
indices=tensor([3, 3]))
29
30
In [*]: a.max(dim=1,keepdim=True)
31
Out[*]:
32
torch.return_types.max(
33
values=tensor([[3.],
34
        [7.]]),
35
indices=tensor([[*],
36
        [*]]))

topk/kth-value

1
In [*]: a
2
Out[*]:
3
tensor([[0., 1., 2., 3.],
4
        [4., 5., 6., 7.]])
5
6
In [*]: a.topk(2,dim=1)
7
Out[*]:
8
torch.return_types.topk(
9
values=tensor([[3., 2.],
10
        [7., 6.]]),
11
indices=tensor([[3, 2],
12
        [3, 2]]))
13
14
In [*]: a.topk(2,dim=1,largest=False)
15
Out[*]:
16
torch.return_types.topk(
17
values=tensor([[0., 1.],
18
        [4., 5.]]),
19
indices=tensor([[0, 1],
20
        [0, 1]]))
21
22
In [*]: a.kthvalue(3,dim=1)
23
Out[*]:
24
torch.return_types.kthvalue(
25
values=tensor([2., 6.]),
26
indices=tensor([2, 2]))

compare

1
In [*]: a
2
Out[*]:
3
tensor([[0., 1., 2., 3.],
4
        [4., 5., 6., 7.]])
5
6
In [*]: a>0
7
Out[*]:
8
tensor([[0, 1, 1, 1],
9
        [1, 1, 1, 1]], dtype=torch.uint8)
10
11
In [*]: torch.gt(a,0)
12
Out[*]:
13
tensor([[0, 1, 1, 1],
14
        [1, 1, 1, 1]], dtype=torch.uint8)
15
16
In [*]: torch.eq(a,a)
17
Out[*]:
18
tensor([[1, 1, 1, 1],
19
        [1, 1, 1, 1]], dtype=torch.uint8)
20
21
In [*]: torch.equal(a,a)
22
Out[*]: True

where / gather

where

根据逻辑值将两个tensor进行聚合

1
In [*]: cond = torch.randn(2,2)
2
3
In [*]: cond
4
Out[*]:
5
tensor([[ 1.9013, -0.6679],
6
        [-1.1109, -1.0334]])
7
8
In [*]: a = torch.ones(2,2)
9
10
In [*]: a
11
Out[*]:
12
tensor([[1., 1.],
13
        [1., 1.]])
14
15
In [*]: b = torch.zeros(2,2)
16
17
In [*]: b
18
Out[*]:
19
tensor([[0., 0.],
20
        [0., 0.]])
21
22
In [*]: torch.where(cond>0,a,b)
23
Out[*]:
24
tensor([[1., 0.],
25
        [0., 0.]])

gather

根据index选择某个tensor中的一部分

1
In [*]: idx = torch.randint(10,[3,3])
2
3
In [*]: idx
4
Out[*]:
5
tensor([[2, 3, 7],
6
        [0, 4, 4],
7
        [6, 4, 4]])
8
9
In [*]: label = (torch.arange(10)+100).expand(3,10)
10
11
In [*]: label
12
Out[*]:
13
tensor([[100, 101, 102, 103, 104, 105, 106, 107, 108, 109],
14
        [100, 101, 102, 103, 104, 105, 106, 107, 108, 109],
15
        [100, 101, 102, 103, 104, 105, 106, 107, 108, 109]])
16
17
In [*]: torch.gather(label,dim=1,index=idx)
18
Out[*]:
19
tensor([[102, 103, 107],
20
        [100, 104, 104],
21
        [106, 104, 104]])
🤣~疯狂暗示~🤣