App下載

pytorch中的矩陣怎么拼接?如何降維?

猿友 2021-07-26 14:43:11 瀏覽數(shù) (4222)
反饋

學(xué)習(xí)機器學(xué)習(xí)必然會學(xué)習(xí)到線性代數(shù)(里面的矩陣),因為不管是numpy還是tensor,又或者是torch,它們的數(shù)據(jù)結(jié)構(gòu)和運算方式都類似于矩陣。然而有些函數(shù)我們是不太熟悉的,今天我們就來介紹一下pytorch矩陣拼接的函數(shù)——cat()和pytorch矩陣降維的函數(shù)——squeeze()吧。

1 squeeze(): 去除size為1的維度,包括行和列。

至于維度大于等于2時,squeeze()不起作用。

行、例:

>>> torch.rand(4, 1, 3)
 
(0 ,.,.) =
  0.5391  0.8523  0.9260
 
(1 ,.,.) =
  0.2507  0.9512  0.6578
 
(2 ,.,.) =
  0.7302  0.3531  0.9442
 
(3 ,.,.) =
  0.2689  0.4367  0.6610
[torch.FloatTensor of size 4x1x3]
>>> torch.rand(4, 1, 3).squeeze()
 
 0.0801  0.4600  0.1799
 0.0236  0.7137  0.6128
 0.0242  0.3847  0.4546
 0.9004  0.5018  0.4021
[torch.FloatTensor of size 4x3]

列、例:

>>> torch.rand(4, 3, 1)
 
(0 ,.,.) =
  0.7013
  0.9818
  0.9723
 
(1 ,.,.) =
  0.9902
  0.8354
  0.3864
 
(2 ,.,.) =
  0.4620
  0.0844
  0.5707
 
(3 ,.,.) =
  0.5722
  0.2494
  0.5815
[torch.FloatTensor of size 4x3x1]
>>> torch.rand(4, 3, 1).squeeze()
 
 0.8784  0.6203  0.8213
 0.7238  0.5447  0.8253
 0.1719  0.7830  0.1046
 0.0233  0.9771  0.2278
[torch.FloatTensor of size 4x3]

不變、例:

>>> torch.rand(4, 3, 2)
 
(0 ,.,.) =
  0.6618  0.1678
  0.3476  0.0329
  0.1865  0.4349
 
(1 ,.,.) =
  0.7588  0.8972
  0.3339  0.8376
  0.6289  0.9456
 
(2 ,.,.) =
  0.1392  0.0320
  0.0033  0.0187
  0.8229  0.0005
 
(3 ,.,.) =
  0.2327  0.6264
  0.4810  0.6642
  0.8625  0.6334
[torch.FloatTensor of size 4x3x2]
>>> torch.rand(4, 3, 2).squeeze()
 
(0 ,.,.) =
  0.0593  0.8910
  0.9779  0.1530
  0.9210  0.2248
 
(1 ,.,.) =
  0.7938  0.9362
  0.1064  0.6630
  0.9321  0.0453
 
(2 ,.,.) =
  0.0189  0.9187
  0.4458  0.9925
  0.9928  0.7895
 
(3 ,.,.) =
  0.5116  0.7253
  0.0132  0.6673
  0.9410  0.8159
[torch.FloatTensor of size 4x3x2]

2 cat函數(shù)

>>> t1=torch.FloatTensor(torch.randn(2,3))
>>> t1
 
-1.9405  1.2009  0.0018
 0.9463  0.4409 -1.9017
[torch.FloatTensor of size 2x3]
>>> t2=torch.FloatTensor(torch.randn(2,2))
>>> t2
 
 0.0942  0.1581
 1.1621  1.2617
[torch.FloatTensor of size 2x2]
>>> torch.cat((t1, t2), 1)
 
-1.9405  1.2009  0.0018  0.0942  0.1581
 0.9463  0.4409 -1.9017  1.1621  1.2617
[torch.FloatTensor of size 2x5]

補充:pytorch中 max()、view()、 squeeze()、 unsqueeze()

查了好多博客都似懂非懂,后來寫了幾個小例子,瞬間一目了然。

一、torch.max()

import torch  
a=torch.randn(3)
print("a:
",a)
print('max(a):',torch.max(a))
 
b=torch.randn(3,4)
print("b:
",b)
print('max(b,0):',torch.max(b,0))
print('max(b,1):',torch.max(b,1))

輸出:

a:
tensor([ 0.9558, 1.1242, 1.9503])
max(a): tensor(1.9503)
b:
tensor([[ 0.2765, 0.0726, -0.7753, 1.5334],
[ 0.0201, -0.0005, 0.2616, -1.1912],
[-0.6225, 0.6477, 0.8259, 0.3526]])
max(b,0): (tensor([ 0.2765, 0.6477, 0.8259, 1.5334]), tensor([ 0, 2, 2, 0]))
max(b,1): (tensor([ 1.5334, 0.2616, 0.8259]), tensor([ 3, 2, 2]))

max(a),用于一維數(shù)據(jù),求出最大值。

max(a,0),計算出數(shù)據(jù)中一列的最大值,并輸出最大值所在的行號。

max(a,1),計算出數(shù)據(jù)中一行的最大值,并輸出最大值所在的列號。

print('max(b,1):',torch.max(b,1)[1])

輸出:只輸出行最大值所在的列號

max(b,1): tensor([ 3,  2,  2])

torch.max(b,1)[0], 只返回最大值的每個數(shù)

二、view()

a.view(i,j)表示將原矩陣轉(zhuǎn)化為i行j列的形式

i為-1表示不限制行數(shù),輸出1列

a=torch.randn(3,4)
print(a)

輸出:

tensor([[-0.8146, -0.6592, 1.5100, 0.7615],
[ 1.3021, 1.8362, -0.3590, 0.3028],
[ 0.0848, 0.7700, 1.0572, 0.6383]])

b=a.view(-1,1)
print(b)

輸出:

tensor([[-0.8146],
[-0.6592],
[ 1.5100],
[ 0.7615],
[ 1.3021],
[ 1.8362],
[-0.3590],
[ 0.3028],
[ 0.0848],
[ 0.7700],
[ 1.0572],
[ 0.6383]])

i為1,j為-1表示不限制列數(shù),輸出1行

b=a.view(1,-1)
print(b)

輸出:

tensor([[-0.8146, -0.6592, 1.5100, 0.7615, 1.3021, 1.8362, -0.3590,
0.3028, 0.0848, 0.7700, 1.0572, 0.6383]])

i為-1,j為2表示不限制行數(shù),輸出2列

b=a.view(-1,2)
print(b)

輸出:

tensor([[-0.8146, -0.6592],
[ 1.5100, 0.7615],
[ 1.3021, 1.8362],
[-0.3590, 0.3028],
[ 0.0848, 0.7700],
[ 1.0572, 0.6383]])

i為-1,j為3表示不限制行數(shù),輸出3列

i為4,j為3表示輸出4行3列

b=a.view(-1,3)
print(b)
b=a.view(4,3)
print(b)

輸出:

tensor([[-0.8146, -0.6592, 1.5100],
[ 0.7615, 1.3021, 1.8362],
[-0.3590, 0.3028, 0.0848],
[ 0.7700, 1.0572, 0.6383]])
tensor([[-0.8146, -0.6592, 1.5100],
[ 0.7615, 1.3021, 1.8362],
[-0.3590, 0.3028, 0.0848],
[ 0.7700, 1.0572, 0.6383]])

三、

1.torch.squeeze()

壓縮矩陣,我理解為降維

a.squeeze(i) 壓縮第i維,如果這一維維數(shù)是1,則這一維可有可無,便可以壓縮

import torch  
a=torch.randn(1,3,4)
print(a)
b=a.squeeze(0)
print(b)
c=a.squeeze(1)
print(c

輸出:

tensor([[[ 0.4627, 1.6447, 0.1320, 2.0946],
[-0.0080, 0.1794, 1.1898, -1.2525],
[ 0.8281, -0.8166, 1.8846, 0.9008]]])

一頁三行4列的矩陣

第0維為1,則可以通過squeeze(0)刪掉,轉(zhuǎn)化為三行4列的矩陣

tensor([[ 0.4627, 1.6447, 0.1320, 2.0946],
[-0.0080, 0.1794, 1.1898, -1.2525],
[ 0.8281, -0.8166, 1.8846, 0.9008]])

第1維不為1,則不可以壓縮

tensor([[[ 0.4627, 1.6447, 0.1320, 2.0946],
[-0.0080, 0.1794, 1.1898, -1.2525],
[ 0.8281, -0.8166, 1.8846, 0.9008]]])

2.torch.unsqueeze()

unsqueeze(i) 表示將第i維設(shè)置為1

對壓縮為3行4列后的矩陣b進行操作,將第0維設(shè)置為1

c=b.unsqueeze(0)
print(c)

輸出一個一頁三行四列的矩陣

tensor([[[ 0.0661, -0.2386, -0.6610, 1.5774],
[ 1.2210, -0.1084, -0.1166, -0.2379],
[-1.0012, -0.4363, 1.0057, -1.5180]]])

將第一維設(shè)置為1

c=b.unsqueeze(1)
print(c)

輸出一個3頁,一行,4列的矩陣

tensor([[[-1.0067, -1.1477, -0.3213, -1.0633]],
[[-2.3976, 0.9857, -0.3462, -0.3648]],
[[ 1.1012, -0.4659, -0.0858, 1.6631]]])

另外,squeeze、unsqueeze操作不改變原矩陣。

以上就是pytorch矩陣拼接和矩陣降維的全部內(nèi)容了,希望能給大家一個參考,也希望大家多多支持W3Cschool。


0 人點贊