Pytoch learning notes -- Summary of common functions of pytoch 1

Catalog ​​​​​​​

1-torch.randn() function  

2-set() Functions and sorted() function

3-DataLoader() Functions and Dataset class

4-.t() function

5- Maximum pooling (max_pool2d) And average pooling (avg_pool2d) function

---

​​​​​​​

1-torch.randn() function  

import torch

batch_size = 1
seq_len = 3
input_size = 4

inputs = torch.randn(seq_len, batch_size, input_size) 

torch.randn() The function is used to generate a group with an average of 0, The variance of 1( Standard normal distribution ) The random number . Examples are as follows ： 

import torch

print(torch.randn(3, 2, 3, 3))

​​​​​​​

torch.randn(seq_len, batch_size, input_size)： The first parameter seq_len Represents sequence length , In the example, the sequence length is 3; The second parameter batch_size Indicates the batch size , The batch size in the example is 2; The third parameter input_size Is the dimension of the input vector , The example is (3, 3）.（ stay RNN Can be understood as ： Example , share 3 A sequence of , Each sequence is divided into 2 batch , The dimension of each batch is 3*3.）

#####################################

#####################################

2-set() Functions and sorted() function

self.country_list = list(sorted(set(self.countries))) # set() duplicate removal , Delete duplicate data ; sorted() Sort 

set() Function is used to delete duplicate data elements ;sorted() Used for sorting elements , Examples are as follows ：

a = ['china', 'china', 'japan']
print(list(set(a)))
print(list(sorted(set(a))))

because ‘c’ < 'j', therefore ‘china’ be ranked at ‘japan’ front .

#####################################

#####################################

3-DataLoader() Functions and Dataset class

from torchvision import datasets
from torch.utils.data import DataLoader, Dataset

batch_size = 64
transform = transforms.Compose([
    transforms.ToTensor(), # take shape by (H, W, C) Of img To shape by (C, H, W) Of tensor, Normalize each value to [0,1]
    transforms.Normalize((0.1307, ), (0.3081, )) # Data standardization by channel 
])

train_dataset = datasets.MNIST(root = '../dataset/mnist/', train = True, download = True, transform = transform)

train_loader = DataLoader(train_dataset, shuffle = True, batch_size = batch_size)

test_dataset = datasets.MNIST(root = '../dataset/mnist/', train = False, download = True, transform = transform)

test_loader = DataLoader(test_dataset, shuffle = False, batch_size = batch_size)

​​​​​​​DataLoader() The data set imported by the function is Dataset type ,shuffle Indicates whether the data set is disturbed .

#####################################

#####################################

4-.t() function

.t() The delta function is going to Tensor Transposition , Examples are as follows ：

import torch

input = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])

print(input)
print(input.t())

#####################################

#####################################

5- Maximum pooling (max_pool2d) And average pooling (avg_pool2d) function

import torch
import torch.nn.functional as F

input = torch.tensor([[[1, 2, 3, 1], [4, 5, 6, 1], [7, 8, 9, 1]]]).unsqueeze(0).float() # unsqueeze(0) In the 0 Add a dimension before the dimension 
print(input.size())
output = F.max_pool2d(input, kernel_size = (1, 4))
print(output)

max_pool2d()： Maximum pooling operation . According to the set core size , Select the largest element value . Example , The nuclear size is (1,4）, It can be understood as picking out the maximum element value of each line .

It should be noted that ：unsqueeze(0) The function of is in the 0 Expand a dimension before the dimension , therefore input Of size by (1, 1, 3, 4).

###

import torch
import torch.nn.functional as F

input = torch.randn(1, 1, 4, 4)
print(input.size())
print(input)
output = F.avg_pool2d(input, kernel_size = (2, 2))
print(output)

 avg_pool2d()： Average pooling operations . According to the set core size , Calculate the average value of the elements in the core .

The role of pooling ： Dimension reduction ;​​​​​​​ Suppress noise , Reduce information redundancy ; Improve the scale invariance of the model 、 Rotation does not deform ; Reduce the amount of model calculation ; Prevent over fitting .