PaddlePaddle/PLSC

PaddlePaddle/PLSC

Description: Paddle Large Scale Classification Tools，supports ArcFace, CosFace, PartialFC, Data Parallel + Model Parallel. Model includes ResNet, ViT, Swin, DeiT, CaiT, FaceViT, MoCo, MAE, ConvMAE, CAE.

Language: Python

License: Apache-2.0

Stars: 155

Forks: 34

Open issues: 22

Created: 2019-12-13T16:30:59Z

Pushed: 2023-06-06T11:13:52Z

Default branch: master

Fork: no

Archived: no

README:

------------------------------------------------------------------------------------------

Introduction

PLSC is an open source repo for a collection of Paddle Large Scale Classification Tools, which supports large-scale classification model pre-training as well as finetune for downstream tasks.

Available Models

• [Face Recognition](./task/recognition/face/)
• [ViT](./task/classification/vit/)
• [Swin](./task/classification/swin/)
• [DeiT](./task/classification/deit/)
• [CaiT](./task/classification/cait/)
• [ConvNeXt](./task/classification/convnext)
• [MoCo v3](./task/ssl/mocov3/)
• [CAE](./task/ssl/cae/)
• [MAE](./task/ssl/mae/)
• [ConvMAE](./task/ssl/mae/)
• [ToMe](./task/accelerate/tome)

Top News 🔥

Update (2023-01-11): PLSC v2.4 is released, we refactored the entire repository based on task types. This repository has been adapted to PaddlePaddle release 2.4. In terms of models, we have added 4 new ones, including FaceViT, CaiT, MoCo v3, MAE. At present, each model in the repository can be trained from scratch to achieve the original official accuracy, especially the training of ViT-Large on the ImageNet21K dataset. In addition, we also provide a method for ImageNet21K data preprocessing. In terms of AMP training, PLSC uses FP16 O2 training by default, which can speed up training while maintaining accuracy.

Update (2022-07-18): PLSC v2.3 is released, a new upgrade, more modular and highly extensible. Support more tasks, such as ViT, DeiT. The static graph mode will no longer be maintained as of this release.

Update (2022-01-11): Supported NHWC data format of FP16 to improve 10% throughtput and decreased 30% GPU memory. It supported 92 million classes on single node 8 NVIDIA V100 (32G) and has high training throughtput. Supported best checkpoint save. And we released 18 pretrained models and PLSC v2.2.

Update (2021-12-11): Released Zhihu Technical Artical and Bilibili Open Class

Update (2021-10-10): Added FP16 training, improved throughtput and optimized GPU memory. It supported 60 million classes on single node 8 NVIDIA V100 (32G) and has high training throughtput.

Update (2021-09-10): This repository supported both static mode and dynamic mode to use paddlepaddle v2.2, which supported 48 million classes on single node 8 NVIDIA V100 (32G). It added PartialFC, SparseMomentum, and ArcFace, CosFace (we refer to MarginLoss). Backbone includes IResNet and MobileNet.

Installation

PLSC provides two usage methods: one is as an external third-party library, and users can use import plsc in their own projects; the other is to develop and use it locally based on this repository.

Note: With the continuous iteration of the PaddlePaddle version, the PLSC master branch adapts to the PaddlePaddle develop branch, and API mismatches may occur in lower versions of PaddlePaddle.

Install plsc as a third-party library

pip install git+https://github.com/PaddlePaddle/PLSC@master

For stable development, you can install a previous version of plsc.

pip install plsc==2.4

Install plsc locally

git clone https://github.com/PaddlePaddle/PLSC.git
cd /path/to/PLSC/
# [optional] pip install -r requirements.txt
python setup.py develop

See [Installation instructions](./tutorials/get_started/installation.md).

Getting Started

See [Quick Run Recognition](./tutorials/get_started/quick_run_recognition.md) for the basic usage of PLSC.

Tutorials

• [Configuration](./tutorials/basic/config.md)

See more [tutorials](./tutorials/README.md).

Documentation

• [Data Augmentation](./docs/data_augmentation.md)

See [documentation](./docs/README.md) for the usage of more APIs or modules.

License

This project is released under the [Apache 2.0 license](./LICENSE).

Citation

@misc{plsc,
title={PLSC: An Easy-to-use and High-Performance Large Scale Classification Tool},
author={PLSC Contributors},
howpublished = {\url{https://github.com/PaddlePaddle/PLSC}},
year={2022}
}