pytorch中的squeeze函数、cat函数使用

1 squeeze(): 去除size为1的维度，包括行和列。

至于维度大于等于2时，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]

不变、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函数

>>> 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]

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

查了好多博客都似懂非懂，后来写了几个小例子，瞬间一目了然。

一、torch.max()

import torch  
a=torch.randn(3)
print("a:\n",a)
print('max(a):',torch.max(a))

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

输出：

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),用于一维数据，求出最大值。

max(a,0),计算出数据中一列的最大值，并输出最大值所在的行号。

max(a,1),计算出数据中一行的最大值，并输出最大值所在的列号。

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

输出：只输出行最大值所在的列号

max(b,1): tensor([ 3,  2,  2])

torch.max(b,1)[0]， 只返回最大值的每个数

二、view()

a.view(i,j)表示将原矩阵转化为i行j列的形式

i为-1表示不限制行数，输出1列

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

输出：

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)

输出：

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为1，j为-1表示不限制列数，输出1行

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

输出：

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为-1，j为2表示不限制行数，输出2列

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

输出：

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为-1，j为3表示不限制行数，输出3列

i为4，j为3表示输出4行3列

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

输出：

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()

压缩矩阵，我理解为降维

a.squeeze(i) 压缩第i维，如果这一维维数是1，则这一维可有可无，便可以压缩

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

输出：

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]]])

一页三行4列的矩阵

第0维为1，则可以通过squeeze(0)删掉，转化为三行4列的矩阵

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维不为1，则不可以压缩

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) 表示将第i维设置为1

对压缩为3行4列后的矩阵b进行操作，将第0维设置为1

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

输出一个一页三行四列的矩阵

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]]])

将第一维设置为1

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

输出一个3页，一行，4列的矩阵

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操作不改变原矩阵

来源：https://blog.csdn.net/abc781cba/article/details/79663190