[pytorch]基础函数笔记

1. torch.randn((dim1, dim2))

import torch
# 1.torch.randn
s = torch.randn((5,2,3)) 

2. torch.linspace(start, end, steps=100, out=None)

# 2.torch.linspace
a = torch.linspace(-1, 1, s.shape[-1]) # 均匀分割。 [-1, 1]分成s.shape[0]
b = torch.linspace(-1, 1, s.shape[-2])
print('a:', a)
print('b:', b)
print('-'* 100)


a: tensor([-1.,  0.,  1.])
b: tensor([-1.,  1.])

3. torch.meshgird(x, y)

# a: tensor([-1.,  0.,  1.])
# b: tensor([-1.,  1.])
# 3.torch.meshgrid
a, b = torch.meshgrid(a, b) # meshgrid后为[3, 2] 但是根据元素的不同 a,b也不同

# 也就是将a，b扩展为 a x b的大小，可用于将一维的坐标扩展为二维上的坐标。（x或者y在变换中其中有一个坐标是保持不变的）
print(a.size(), b.size())
print('meshgrid a:', a)
print('meshgrid b:', b)
print('-'* 100)

torch.Size([3, 2]) torch.Size([3, 2])
meshgrid a: tensor([[-1., -1.],
        [ 0.,  0.],
        [ 1.,  1.]])
meshgrid b: tensor([[-1.,  1.],
        [-1.,  1.],
        [-1.,  1.]])

4. torch.expand(*sizes) → Tensor

# 4.torch.expand
a = a.expand(s.shape[0], 1, -1, -1) # 第三第四个参数为-1,-1 表示使用a.shape(3, 2)本身的2个维度 
b = b.expand(s.shape[0], 1, -1, -1) # [5, 1, 3, 2] 上同
print(a.size())
print(b.size())


torch.Size([5, 1, 3, 2])
torch.Size([5, 1, 3, 2])

5.torch.cat(tensors,dim=0,out=None)→ Tensor 和 torch.stack(tensors,dim=0,out=None)→ Tensor

# 5.torch.cat()
cat_ab = torch.cat((a,b), dim=1)  # 使用cat的前提是 其他维度必须相同。对dim=1上进行cat 
print("cat_ab:", cat_ab.size()) # [5, 2, 3, 2]

>>> a = torch.rand((2, 3))
>>> b = torch.rand((2, 3))
>>> c = torch.cat((a, b))
>>> a.size(), b.size(), c.size()
(torch.Size([2, 3]), torch.Size([2, 3]), torch.Size([4, 3]))

>>> a = torch.rand((2, 3))
>>> b = torch.rand((2, 3))
>>> c = torch.stack((a, b))
>>> a.size(), b.size(), c.size()
(torch.Size([2, 3]), torch.Size([2, 3]), torch.Size([2, 2, 3]))

6. torch.nn.functional.interpolate(input_shape, scale_factor= , mode='')

# 6.torch.nn.functional.interpolate
cat_ab = torch.nn.functional.interpolate(cat_ab, scale_factor=2, mode="bilinear") # [N, C, H, W] --> [N, C, 2H, 2W]
print("interpolate cat_ab:", cat_ab.size()) # [5, 2, 6, 4] 

cat_ab: torch.Size([5, 2, 3, 2])
interpolate cat_ab: torch.Size([5, 2, 6, 4])

7. torch.sum() , torch.ones()

import torch
a = torch.ones((3,2)) #  torch.Size([3, 2])
b = a.sum(dim=0)
b.size() #  torch.Size([2])

c = a.sum(dim=1)
b.size() #  torch.Size([3])

8. torch.nonzero() 返回查找符合条件的索引

(x condition y) .nonzero()

x=torch.arange(12).view(3,4)
x
tensor([[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11]])
(x>4).nonzero()
tensor([[1, 1],
        [1, 2],
        [1, 3],
        [2, 0],
        [2, 1],
        [2, 2],
        [2, 3]])

9. torch.clamp(input, min, max, *, out=None) → Tensor

>>> a = torch.randn(4)
>>> a
tensor([-1.7120,  0.1734, -0.0478, -0.0922])
>>> torch.clamp(a, min=-0.5, max=0.5)
tensor([-0.5000,  0.1734, -0.0478, -0.0922])

>>> a = torch.randn(4)
>>> a
tensor([-0.0299, -2.3184,  2.1593, -0.8883])
>>> torch.clamp(a, min=0.5)
tensor([ 0.5000,  0.5000,  2.1593,  0.5000])

>>> a = torch.randn(4)
>>> a
tensor([ 0.7753, -0.4702, -0.4599,  1.1899])
>>> torch.clamp(a, max=0.5)
tensor([ 0.5000, -0.4702, -0.4599,  0.5000])

10. torch.squeeze(input, dim) and torch.unsqueeze(input, dim)

增加或者减少某维度为1的维数

>>> x = torch.zeros(2, 1, 2, 1, 2)
>>> x.size()
torch.Size([2, 1, 2, 1, 2])
>>> y = torch.squeeze(x)
>>> y.size()
torch.Size([2, 2, 2])
>>> y = torch.squeeze(x, 0)
>>> y.size()
torch.Size([2, 1, 2, 1, 2])
>>> y = torch.squeeze(x, 1)
>>> y.size()
torch.Size([2, 2, 1, 2])

>>> x = torch.tensor([1, 2, 3, 4])
>>> torch.unsqueeze(x, 0)
tensor([[ 1,  2,  3,  4]])
>>> torch.unsqueeze(x, 1)
tensor([[ 1],
        [ 2],
        [ 3],
        [ 4]])

11. torch.nn.functional.max_pool2d 以及 关于tensor之间的比较，bool值的转化

import torch
import torch.nn as nn
heat = torch.tensor([[[[-0.9752, -2.1402],
          [-0.67, -1.0918],
          [ -1.5324,  0.7099]],

         [[-0.3685, -0.1001],
          [-0.8646, -1.4367],
          [ 0.9132, -1.2989]]]])
kernel = 2
ans = nn.functional.max_pool2d(
        heat, (kernel, kernel), stride=1, padding=1)
ans.size()
#1.
t1
'''
tensor([[[[-0.9752, -0.9752],
          [-0.6700, -0.6700],
          [-0.6700,  0.7099]],

         [[-0.3685, -0.1001],
          [-0.3685, -0.1001],
          [ 0.9132,  0.9132]]]])
'''
t1 = (ans[:, :, :-1, :-1] == heat)
'''
tensor([[[[ True, False],
          [ True, False],
          [False,  True]],

         [[ True,  True],
          [False, False],
          [ True, False]]]])
'''

#2.
t2 =(ans[:, :, :-1, :-1] ==  heat).float()
'''
tensor([[[[1., 0.],
          [1., 0.],
          [0., 1.]],

         [[1., 1.],
          [0., 0.],
          [1., 0.]]]])
'''

关于条件判断后，通过索引查找

a = torch.tensor([[2, 1, 0], [1, 2, 2]])
ind  =  (a > 0.1)
'''
tensor([[ True,  True, False],
        [ True,  True,  True]])
'''     
b = torch.ones((2,3))
c = b[ind]
 
'''
tensor([1., 1., 1., 1.])
'''

12. torch.Tensor.view(*shape) → Tensor

>>> x = torch.randn(4, 4)
>>> x.size()
torch.Size([4, 4])
>>> y = x.view(16)
>>> y.size()
torch.Size([16])
>>> z = x.view(-1, 8)  # the size -1 is inferred from other dimensions
>>> z.size()
torch.Size([2, 8])

>>> a = torch.randn(1, 2, 3, 4)
>>> a.size()
torch.Size([1, 2, 3, 4])
>>> b = a.transpose(1, 2)  # Swaps 2nd and 3rd dimension
>>> b.size()
torch.Size([1, 3, 2, 4])
>>> c = a.view(1, 3, 2, 4)  # Does not change tensor layout in memory
>>> c.size()
torch.Size([1, 3, 2, 4])
>>> torch.equal(b, c)
False

13. torch.cumsum(input, dim, *, dtype=None, out=None) → Tensor

a = torch.randn((5,2))
a
tensor([[-0.9153,  0.7133],
        [ 0.0106,  1.4720],
        [ 0.5626, -0.4366],
        [ 0.3931, -0.4698],
        [-0.5168,  0.0720]])
a.cumsum(0)
tensor([[-0.9153,  0.7133],
        [-0.9047,  2.1853],
        [-0.3421,  1.7487],
        [ 0.0510,  1.2789],
        [-0.4658,  1.3510]])
a.cumsum(1)
tensor([[-0.9153, -0.2020],
        [ 0.0106,  1.4827],
        [ 0.5626,  0.1259],
        [ 0.3931, -0.0766],
        [-0.5168, -0.4448]])

14. torch.argsort(input, dim=-1, descending=False) → LongTensor

Example:

>>> a = torch.randn(4, 4)
>>> a
tensor([[ 0.0785,  1.5267, -0.8521,  0.4065],
        [ 0.1598,  0.0788, -0.0745, -1.2700],
        [ 1.2208,  1.0722, -0.7064,  1.2564],
        [ 0.0669, -0.2318, -0.8229, -0.9280]])


>>> torch.argsort(a, dim=1)
tensor([[2, 0, 3, 1],
        [3, 2, 1, 0],
        [2, 1, 0, 3],
        [3, 2, 1, 0]])

14. torch.mm(input, mat2, *, out=None) → Tensor

>>> mat1 = torch.randn(2, 3)
>>> mat2 = torch.randn(3, 3)
>>> torch.mm(mat1, mat2)
tensor([[ 0.4851,  0.5037, -0.3633],
        [-0.0760, -3.6705,  2.4784]])

15. torch.triu(input, diagonal=0, *, out=None) → Tensor

>>> a = torch.randn(4, 4)
>>> a
tensor([[ 0.5144,  0.5091, -0.3698,  0.3694],
        [-1.1344, -0.2793,  1.6651, -1.3632],
        [-0.3397, -0.1468, -0.0300, -1.1186],
        [-2.1449,  1.3087, -0.1409,  2.4678]])
>>>
>>> torch.triu(a)
tensor([[ 0.5144,  0.5091, -0.3698,  0.3694],
        [ 0.0000, -0.2793,  1.6651, -1.3632],
        [ 0.0000,  0.0000, -0.0300, -1.1186],
        [ 0.0000,  0.0000,  0.0000,  2.4678]])
>>>
>>> torch.triu(a, diagonal=1)
tensor([[ 0.0000,  0.5091, -0.3698,  0.3694],
        [ 0.0000,  0.0000,  1.6651, -1.3632],
        [ 0.0000,  0.0000,  0.0000, -1.1186],
        [ 0.0000,  0.0000,  0.0000,  0.0000]])
>>>
>>> torch.triu(a, diagonal=-1)
tensor([[ 0.5144,  0.5091, -0.3698,  0.3694],
        [-1.1344, -0.2793,  1.6651, -1.3632],
        [ 0.0000, -0.1468, -0.0300, -1.1186],
        [ 0.0000,  0.0000, -0.1409,  2.4678]])
>>>
>>> torch.triu(a, diagonal=-2)
tensor([[ 0.5144,  0.5091, -0.3698,  0.3694],
        [-1.1344, -0.2793,  1.6651, -1.3632],
        [-0.3397, -0.1468, -0.0300, -1.1186],
        [ 0.0000,  1.3087, -0.1409,  2.4678]])
>>>
>>> torch.triu(a, diagonal=-3)
tensor([[ 0.5144,  0.5091, -0.3698,  0.3694],
        [-1.1344, -0.2793,  1.6651, -1.3632],
        [-0.3397, -0.1468, -0.0300, -1.1186],
        [-2.1449,  1.3087, -0.1409,  2.4678]])

16. torch.nn.functional.interpolate(input, size=None, scale_factor=None, mode='nearest', align_corners=None, recompute_scale_factor=None)

ins_feat = F.interpolate(ins_feat, size=(self.seg_num_grids[idx] * sa_h, self.seg_num_grids[idx] * sa_w), mode='bilinear')

17. torch.zeros(*size, *, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

>>> torch.zeros(2, 3)
tensor([[ 0.,  0.,  0.],
        [ 0.,  0.,  0.]])

>>> torch.zeros(5)
tensor([ 0.,  0.,  0.,  0.,  0.])

18. pytorch通过索引直接扩大矩阵

a = torch.randn((4,5,5))
b = torch.ones(10,dtype=int) #注意这里必须是bool,long或者byte,否则无法构成索引

c = a[b] 
c.size()

------------------------------------------------------------------------
OUTPUTS:
tensor([[[-1.0732, -0.7315, -0.4887,  1.3547, -0.8984],
         [-1.3814,  0.7787,  0.4166, -0.4355, -0.4207],
         [-0.1047,  2.1872,  0.3320, -0.4274, -0.2006],
         [-0.1602,  0.1737,  1.0411,  0.8609,  0.3690],
         [ 0.2870, -0.4992, -0.4487,  1.0585,  0.2058]],

        [[ 0.4890, -0.0888, -0.2893, -0.3513,  2.1535],
         [ 0.2390, -1.3600,  0.2401,  1.2936, -1.7246],
         [ 0.3906,  0.9688, -0.1664,  1.9916, -0.5115],
         [ 0.4018,  0.0544, -1.4237, -0.3007,  0.1592],
         [ 1.2119,  1.0150,  2.2753,  0.2863, -2.0256]],

        [[ 0.1236,  0.5212, -0.2304, -0.0847,  0.7568],
         [ 0.3000, -0.2914, -0.9238,  0.2331, -0.3673],
         [ 0.0475, -0.8758,  0.8781,  0.4415,  0.2434],
         [ 0.4388, -2.3196,  1.0323,  0.8560, -1.8282],
         [-2.3422, -0.0589,  0.7717, -0.1974,  0.2317]],

        [[ 1.5203, -0.4977, -1.0504, -0.0240, -0.1234],
         [ 0.4833, -0.4155,  1.3780,  0.2443,  0.8457],
         [-0.4753, -0.3182,  0.1281, -0.1161,  0.8857],
         [ 1.8742,  0.5412, -0.8276,  1.2642,  0.3902],
         [ 2.4383,  0.9594,  0.0306, -0.1631,  0.8616]]])
------------------------------------------------------------------------
b:
tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1])
------------------------------------------------------------------------
c
tensor([[[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]],

        [[ 2.6001,  1.7210,  0.2823,  0.0699, -0.2888],
         [-0.4970,  0.2690, -0.2260,  0.3258,  0.4532],
         [ 0.2565, -1.2199,  0.1472, -1.5703,  1.5488],
         [-1.2409, -0.1028, -1.0456, -0.7941, -0.4220],
         [-1.1634, -1.8762,  0.0277, -0.0281, -1.9816]]])

19. torch.repeat(*sizes) → Tensor

>>> x = torch.tensor([1, 2, 3])
>>> x.repeat(4, 2) #根据x的维度 x[1, 3] 变换 1 * 4 = 4,  3 * 2 = 6  [4, 6] 
tensor([[ 1,  2,  3,  1,  2,  3],
        [ 1,  2,  3,  1,  2,  3],
        [ 1,  2,  3,  1,  2,  3],
        [ 1,  2,  3,  1,  2,  3]])
>>> x.repeat(4, 2, 1).size()# 相当于x的看成[1, 1, 3]
torch.Size([4, 2, 3])

20. 广播以及在pdb调试中画掩码

# a[5 , 1]  b[1, 2]  (a-b).size() --> (5,2) 我认为这是一种广播！
 a = torch.randn((5,1)) #最后一维为1
 b = torch.randn((1,10))# 第一维度为1
 print(a)
 print(b)
 c = a - b
 print(c.size())
 print(c)

tensor([[-0.7045],
        [ 1.2742],
        [ 0.4087],
        [ 0.4675],
        [ 0.2742]])
tensor([[ 0.0117,  1.2361,  0.8163,  0.1032,  1.3831,  1.3738,  1.4786, -0.4542,
          0.1732,  3.4909]])
torch.Size([5, 10])
tensor([[-0.7162, -1.9407, -1.5209, -0.8078, -2.0877, -2.0783, -2.1831, -0.2504,
         -0.8778, -4.1955],
        [ 1.2625,  0.0380,  0.4578,  1.1709, -0.1090, -0.0996, -0.2044,  1.7283,
          1.1009, -2.2168],
        [ 0.3971, -0.8274, -0.4076,  0.3055, -0.9744, -0.9651, -1.0698,  0.8629,
          0.2355, -3.0822],
        [ 0.4558, -0.7687, -0.3488,  0.3643, -0.9156, -0.9063, -1.0111,  0.9216,
          0.2943, -3.0235],
        [ 0.2625, -0.9620, -0.5422,  0.1709, -1.1089, -1.0996, -1.2044,  0.7283,
          0.1010, -3.2168]])

# 这样也可以
a = torch.randn((2,5,1))
b = torch.randn((2,1,10))
c = a - b
#
a = torch.randn((5,1))
b = torch.randn((5,10))

# 这样就会报错
a = torch.randn((5,2))
b = torch.randn((1,10))
# 
a = torch.randn((5,1))
b = torch.randn((2,10))
 #
a = torch.randn((5,1))
b = torch.randn((10,1))

# 画图
from PIL import Image
import numpy as np
img = Image.fromarray(np.array(gt_masks.cpu()*255).astype(‘uint8’)).convert(‘RGB’)
img.save(‘picture.jpg’)

21. torch.nn.CrossEntropyLoss()

易错点： 如果input是二维的（N，C），那么target就是（N）；如果input是三维的（N,C,d_k)，那么target就是(N, d_k) ，而不是(N,C)！！

# 三维
import torch
import torch.nn as nn
loss = nn.CrossEntropyLoss()
input = torch.randn(2,92, 100, requires_grad=True)
target = torch.empty(2,100, dtype=torch.long).random_(5)
print(input, target)
output = loss(input, target)


# 二维
import torch
import torch.nn as nn
loss = nn.CrossEntropyLoss()
input = torch.randn(92, 100, requires_grad=True)
# input = torch.randn(100, 92, requires_grad=True) 则会出错
target = torch.empty(92, dtype=torch.long).random_(5)
print(input, target)
output = loss(input, target)