whu-pzhang / pytorch_styleguide

📖 English    |    📖 中文

A PyTorch style-guide & best practices

The offical PyTorch Tutorials is the best materials for beginers.

recommend using Python3.6

From our experience, we recommend using Python 3.6 because of the compatibility and several new features which became very handy for clean and simple code.

File Organization

Don't put all stuffs into the same file! A best practices is to separate the model into a separate file, such as models/unet.py. And keep the layers, losses in respective files(losses/unet_loss.py). The finally models contained multiple networkds should be reference in a file with its name(e.g, yolov3.py, GCGAN.py).

The main routine, such as the train and test scripts should only import from the file having the model's name.

├── README.md
├── config
├── losses
├── models
├── requirements.txt
├── test
├── train
└── utils

• README.md: descriptions of your project
• requirements.txt: the required packages of your project
• config: configuration file folder. such as your training parameters.
• models: contain the basic backbone architectures and final models.
• losses: losses which not contained in native PyTorch, such as yolov3_loss.
• utils: other tools for data prepare and post processing, such as split large remote sensing image into your network desired size.
• train: as your can known from the folder name. (cloud_train.py, road_train.py ...)
• test: same as above. (cloud_detection.py, road_detection.py ...)

Naming Conventions

Follow the Google Styleguide for Python.

Here is a summary of the most commonly used rules:

Type	Convention	Example
Packages & Modules	lower_with_under	from prefetch_generator import BackgroundGenerator
Classes	CapWords	class DataLoader
Constants	CAPS_WITH_UNDER	BATCH_SIZE=16
Instances	lower_with_under	dataset = Dataset
Methods & Functions	lower_with_under()	def visualize_tensor()
Variables	lower_with_under	background_color='Blue'

Try your best to using meaningful words.

Demo

MNIST digits recognition

$ python train/mnist_train.py --cfg config/mnist.yaml

TODO

• U-Net