How to standardize the creation of a pytorch project

A good Pytorch The project shall include the following specifications ：

Definition PyTorch Software engineering specification of the project , contain ：

• Code directory for training tests / Document specification ;
• Configuration specification ;

Code directory for training tests / Document specification

Fit every deepvac canonical PyTorch Model training program , Include the following directories and files ：

Catalog / file	explain	Whether must
README.md	Project description 、git Number of branches and their introduction 、 Description of the storage path of the original data	yes
train.py	Model training and validation documentation , Inherit DeepvacTrain Class extension implementation	yes
test.py	Model test file , Inherit Deepvac Class extension implementation	yes
config.py	Project overall configuration file	yes
modules/model.py	Model definition file ,PyTorch Module Class extension implementation	In the case of a single model , yes
modules/model_{name}.py	ditto , There are many. model When , Use suffix distinguish	In the case of multiple models , yes
modules/loss.py	loss Realization . If the implementation is lightweight , It can be placed directly in modules/model.py in	no
modules/utils.py	Tool class / Method definition file	no
modules/utils_{name}.py	ditto , There are multiple tool classes / Function file , Use suffix distinguish	no
synthesis/synthesis.py	Data synthesis or cleaning code	no
synthesis/config.py	synthesis.py Configuration file for	no
data/dataloader.py	dataset Class	no
data/train.txt	Training set manifest file	no
data/val.txt	Validation set manifest file	no
aug/aug.py	Data enhanced code . If the implementation is lightweight , It can be placed directly in dataset Class	no
aug/config.py	aug.py Configuration file for	no
log/*.log	Log output directory	yes
output/model__*.pth	Export or import model file	Default Deepvac Output
output/checkpoint__*.pth	Output or input checkpoint file	Default Deepvac Output

These files cover a PyTorch The whole life cycle of model training ：

• Raw data , stay README.md Description in ;
• Data cleaning / synthesis , stay synthesis/synthesis.py In the definition of ;
• Data to enhance , stay aug/aug.py In the definition of （ Light weight can be achieved in dataset Definition in class ）;
• data input , stay data/dataloader.py In the definition of ;
• model training , stay train.py In the definition of ;
• Model validation , stay train.py In the definition of ;
• Model test , stay test.py In the definition of ;
• Model output , stay output To store in a directory ;
• Log output , stay log To store in a directory ;

Configuration specification

• config.py Is a first-class citizen in the norms ;
• The configuration of user interface level is in config.py in ;
• The configurations defined by internal development are all in the class auditConfig In the method , And can be config.py The values in override ;
• When starting distributed training , because –rank and –gpu The parameter is process level , from argparse.ArgumentParser Module to pass , The user needs to specify... On the command line ;
• The input parameters of the constructor of a class are generally config example ;
• Again ,config.py Be a first-class citizen in the norm .

Excerpt from ：DeepVac