Torch learning summary

Catalog

1. torch.clamp()

2.  torch.gather()

3.  torch.bmm

4. torch.squeeze And torch.unsqueeze

5. torch.expand

---

1. torch.clamp()

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

Enter input The tensor of each element is clamped to the interval  [min,max][min,max], And return the result to a new tensor .

The operation is defined as follows ：

         | min, if x_i < min
y_i = | x_i, if min <= x_i <= max
         | max, if x_i > max

  Example code

import torch

a=torch.randint(low=0,high=10,size=(10,1))
print(a)
a=torch.clamp(a,3,6)
print(a)

Running results

tensor([[9],
        [0],
        [4],
        [7],
        [8],
        [0],
        [7],
        [9],
        [2],
        [4]])
tensor([[6],
        [3],
        [4],
        [6],
        [6],
        [3],
        [6],
        [6],
        [3],
        [4]])

2.  torch.gather()

effect ： Collect the value of the specified position of the specific dimension entered

Example code

import torch
input = [
    [2, 3, 4, 5, 0, 0],
    [1, 4, 3, 0, 0, 0],
    [4, 2, 2, 5, 7, 0],
    [1, 0, 0, 0, 0, 0]
]
input = torch.tensor(input)
# Be careful index The type of 
length = torch.LongTensor([[4],[3],[5],[1]])
#index The reason is that 1, Because the sequence dimension is from 0 Start calculated 
out = torch.gather(input, 1, length-1)
prin(out)

  Running results

tensor([[5],
        [3],
        [7],
        [1]])

3.  torch.bmm

Calculate two tensor Matrix multiplication of ,torch.bmm(a,b),tensor a Of size by (b,h,w),tensor b Of size by (b,w,h), Pay attention to two tensor The dimension of must be 3.

Example code

import torch
c = cc=torch.randn((2,2,5))
print(c)
d = torch.reshape(cc,(2,5,2))
print(d)
e = torch.bmm(c, d)
print(e)

Running results

tensor([[[-1.9362,  0.9502,  0.4412,  0.2745, -1.0118],
         [ 0.0709,  1.4860,  0.9381, -0.6637,  1.0663]],

        [[-0.8230, -0.6033,  0.4908,  0.6017,  0.7760],
         [ 0.7727, -0.4746, -0.3565, -0.5635,  1.3207]]])
tensor([[[-1.9362,  0.9502],
         [ 0.4412,  0.2745],
         [-1.0118,  0.0709],
         [ 1.4860,  0.9381],
         [-0.6637,  1.0663]],

        [[-0.8230, -0.6033],
         [ 0.4908,  0.6017],
         [ 0.7760,  0.7727],
         [-0.4746, -0.3565],
         [-0.5635,  1.3207]]])
tensor([[[ 4.8012, -2.3691],
         [-2.1249,  1.0561]],

        [[ 0.0393,  1.3231],
         [-1.6222,  0.9178]]])

Process finished with exit code 0

4. torch.squeeze And torch.unsqueeze

torch.squeeze() This function mainly compresses the dimensions of data , Remove the dimension as 1 Dimension of , For example, a row or a column , One row, three columns （1,3） After removing the first dimension of one, the number becomes （3） That's ok .squeeze(a) Will be a All of them are 1 Delete the dimension of . Not for 1 The dimension of has no effect .a.squeeze(N) Just get rid of a The dimension specified in is one . Another form is b=torch.squeeze(a,N) a Remove the dimension whose specified fixed dimension is one .

torch.unsqueeze() This function is mainly used to expand the data dimension . Add a dimension of one to the specified position , For example, there was originally three lines of data （3）, stay 0 When you add one dimension to the position of, it becomes a row and three columns （1,3）.a.squeeze(N) Is in the a In N Add a dimension of 1 Dimensions . Another form is b=torch.squeeze(a,N) a Is in the a In N Add a dimension of 1 Dimensions
 

Example code

import torch
 
x = torch.zeros(3,2,4,1,2,1)# dimension of 3*2*4*1*2
print(x.size())             # torch.Size([3, 2, 4, 1, 2, 1])
print(x.shape)
 
y = torch.squeeze(x)        # Returns a tensor with all the dimensions of input of size 1 removed.
print(y.size())             # torch.Size([3, 2, 4, 2])
print(y.shape)
 
z = torch.unsqueeze(y,dim=0)# Add a dimension of 1 in the 0th position
print(z.size())             # torch.Size([1, 3, 2, 4, 2])
print(z.shape)
 
z = torch.unsqueeze(y,dim=1)# Add a dimension of 1 in the 1st position
print(z.size())             # torch.Size([3, 1, 2, 4, 2])
print(z.shape)
 
z = torch.unsqueeze(y,dim=2)# Add a dimension of 1 in the 2nd position
print(z.size())             # torch.Size([3, 2, 1, 4, 2])
print(z.shape)
 
y = torch.squeeze(x,dim=0)  # remove the 0th position of 1 (no 1)
print('dim=0', y.size())    # torch.Size([3, 2, 4, 1, 2, 1])
print('dim=0', y.shape)
 
y = torch.squeeze(x, dim=1)  # remove the 1st position of 1 (no 1)
print('dim=1', y.size())     # torch.Size([3, 2, 4, 1, 2, 1])
print('dim=1', y.shape)
 
y = torch.squeeze(x, dim=2)  # remove the 2nd position of 1 (no 1)
print('dim=2', y.size())     # torch.Size([3, 2, 4, 1, 2])
print('dim=2', y.shape)
 
y = torch.squeeze(x, dim=3)  # remove the 3rd position of 1 (yes)
print('dim=3', y.size())     # torch.Size([3, 2, 4, 2])
print('dim=3', y.shape)
 
y = torch.squeeze(x, dim=4)  # remove the 4th position of 1 (no 1)
print('dim=4', y.size())     # torch.Size([3, 2, 4, 1, 2, 1])
print('dim=4', y.shape)
 
y = torch.squeeze(x, dim=5)  # remove the 5th position of 1 (yes)
print('dim=5', y.size())     # torch.Size([3, 2, 4, 1, 2])
print('dim=5', y.shape)
 
y = torch.squeeze(x, dim=6)  # RuntimeError: Dimension out of range (expected to be in range of [-6, 5], but got 6)
print('dim=6', y.size())
print('dim=6', y.shape)
 

5. torch.expand

return tensor A new view of , A single dimension expands to a larger size . tensor It can also be expanded to higher dimensions , New dimensions will be attached to the front . expand tensor There is no need to allocate new memory , Just create a new one tensor The view of , Among them, through stride Set to 0, One dimension will be extended to higher dimensions . Any one-dimensional can be extended to any number without allocating new memory .

Example code

import torch
x = torch.Tensor([[1], [2], [3]])
print(x.size())
y = x.expand(3, 4)

print(x.size())
print(y.size())

print(x)
print(y)

result

torch.Size([3, 1])
torch.Size([3, 1])
torch.Size([3, 4])
tensor([[1.],
        [2.],
        [3.]])
tensor([[1., 1., 1., 1.],
        [2., 2., 2., 2.],
        [3., 3., 3., 3.]])