English | Simplified Chinese

PLSC: PaddlePaddle Large Scale Classification Package

Introduction

The fully-connected layer is usually used as the last layer of deep neural networks for multi-class classification tasks. As the memory occupied by parameters of the fully-connected layer increasing linearly with the increase in number of classes, the training of those kind of neural networks face the following challenges when tasks have a large number of classes:

• GPU memory limitation

• Heavy communication overhead

PaddlePaddle Large Scale Classification (PLSC) package is an end-to-end solution for large scale classification tasks based on PaddlePaddle, which provides a total solution from model training to deployment for users.

Highlights

• Support large scale classification problems：PLSC can support 2.5x more classes using 8 V100 GPU cards, and the maximum number of classes supported increases as the increase in number of GPU cards used;
• High training speed：With one million classes, the training speed is 2,122 images/sec using ResNet50 model and 8 V100 GPU cards are used, and PLSC supports mixed precision training and distributed training;
• Support the adjustment of number of GPU cards used for training：A training starts from checkpoints can use a different number of GPU cards from that used for the last training;
• Builtin pre-precessing tools for base64 format images：PLSC provides builtin pre-processing tools for images in base64 format，including global shuffle, dataset spliting, and so on;
• Support for custom models：PLSC contains the builtin ResNet50, ResNet101 and ResNet152 models，and users can define their customized models；
• Support the automatic uploading and downloading of checkpoints for HDFS file systems;
• A end-to-end solution：PLSC provides a total solution from model training to deployment for users.

Quick Start

Refer to quick start for how to install and use PLSC.

Serving and Deployment

Refer to deployment guider for how to use and deployment serving service.

Advanced Usage

Refer to advanced guider for advanced usage.

API Reference

Refer to API reference for all APIs.

Pre-trained model and performance

Pre-trained model

We provide the following pre-trained model for users to fine-tuning their models.

Model	Description
resnet50_distarcface_ms1m_arcface	This model is trained with ResNet50 and MS1M-ArcFace and the loss type used is dist_arcface. The validation accuracy on lfw dataset is 0.99817.

Performance

Model	Training Dataset	lfw	agendb_30	cfp_ff	cfp_fp
ResNet50	MS1M-ArcFace	0.99817	0.99827	0.99857	0.96314
ResNet50	CASIA	0.9895	0.9095	0.99057	0.915

Note：The loss_type used is 'dist_arcface'. For more information about ArcFace please refer to ArcFace: Additive Angular Margin Loss for Deep Face Recognition.