Introduction to pytorch ecology

        PyTorch The power of is not limited to its own ease of use , What's more, the open source community revolves around PyTorch A series of toolkits （ It's usually Python package） And procedure , These kits are a great convenience PyTorch Use in specific areas . For example, for image processing, we usually use torchvision, For video processing, I will use torchvideo, For text processing, we will use torchtext wait . Together, these excellent toolkits make up  PyTorch The ecology of . Use... In a specific field PyTorch Toolkit in Ecology , It can greatly reduce the entry threshold , It is convenient for novices to quickly master the coding in specific fields .

（ One ）torchvision

        torchvision yes  PyTorch A library for image processing in , It mainly deals with models and data sets in the direction of computer vision .

1. torchvision.datasets： Used to load some built-in data sets and some commonly used functions .
2. torchvision.models： Used to load PyTorch Built in model , And the pre training parameters of these models can be downloaded together .
3. torchvision.transforms： For image transformation , Enhance the data . for example ： Random scaling （transforms.Resize）、 Random position clipping （transforms.RandomCrop）、 Random rotation （transforms.RandomRotation（）） etc. .
4. torchvision.utils： Provides some visual methods , You can splice several pictures together 、 Visualize the effect of detection and segmentation .

import torchvision
from torchvision import datasets, transforms

# Load the built-in model Resnet18
model=torchvision.models.resnet18(pretrained=True)

# Load the built-in dataset MINIST
train_ds=datasets.MNIST('data', train=True, transform=ToTensor(), download=True)
test_ds=datasets.MNIST('data', train=False, transform=ToTensor(), download=True)

# use transforms Enhance data 
transform_train=transforms.Compose([transforms.Resize(224),  # Random scaling 
                                    transforms.RandomCrop(192), # Random position clipping 
                                    transforms.RandomHorizontalFlip(), # Random horizontal flip 
                                    transforms.RandomRotation(0.2),  # Random rotation 
                                    transforms.ColorJitter(brightness=0.5,), # Adjust the brightness of the picture 
                                    transforms.ColorJitter(contrast=0.5,),   # Image data enhancement 
                                    transforms.ToTensor(),# Convert the image into tensor; Change the picture format ; Normalize the pixel value range to （0,1）
                                    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),# Data normalization 

（ Two ）torchcideo

        TorchVideo It is a deep learning library focusing on video understanding .torchVideo It provides the reusable software needed to accelerate the research of video understanding 、 Modular and efficient components .TorchVideo It's using PyTorch Developed , Support different in-depth learning video components , Such as video model 、 Video dataset and video specific conversion .

        PytorchVideo It provides a modular and efficient method to accelerate the research of video understanding API. It also supports different deep learning video components , Such as video model 、 Video dataset and video specific conversion , most important of all ,PytorchVideo Also provided model zoo, So that people can use various advanced pre training video models and their evaluation benchmarks .

（ 3、 ... and ）torchtext

        torchtext yes PyTorch Officially used for natural language processing （NLP） The toolkit .

        torchtext It is convenient to preprocess the text , For example, truncation and lengthening 、 Construct thesaurus, etc .torchtext It mainly includes the following main components ：1. Data processing tools （torchtext.data.functional、torchtext.data.utils）、2. load NLP Built in datasets （torchtext.data.datasets）、3. Building a vocabulary （torchtext.vocab） etc. .

import torch
import torchtext

# Load text built-in dataset 
train_iter, test_iter=torchtext.datasets.IMDB()

#torchtext Built in word segmentation tool 
from torchtext.data.utils import  get_tokenizer
tokenizer=get_tokenizer('basic_english') #basic_english： Will turn letters into lowercase and divide by spaces , Can split punctuation marks .

#torchtext Built in tool for creating thesaurus 
from torchtext.vocab import build_vocab_from_iterator
def yield_tokens(data):   # When creating a thesaurus, you need to pass in a generator , So first define a generator function 
    for (_,text) in data:
        yield tokenizer(text)
vocab=build_vocab_from_iterator(yield_tokens(train_data),
                                specials=['<pad>', '<unk>'],
                                min_freq=3)
vocab.set_default_index(vocab['<unk>'])

notes ：torchtext When installing, its version needs to be the same as that installed by yourself PyTorch The version corresponds to , Otherwise, it will not work properly . You can refer to the following website （https://pypi.org/project/torchtext/）.