This repo is the official implementation of "Large-batch Optimization for Dense Visual Predictions (NeurIPS 2022)". Since we adopted private frameworks (POD and LinkLink) to conduct the experiments previously, the results open-sourced with mmdetection will be slightly different from the results in our paper. The optimized version of DDP will be released in the future.
Step 0. Please refer to mmdetection get started for installation and dataset preparation.
Step 1. Install AGVM from source:
git clone https://github.com/Sense-X/AGVM.git
cd AGVM
make installPlease refer to this doc for examples of training.
The box mAP of Faster R-CNN:
| Batch Size | 32 | 256 | 512 |
|---|---|---|---|
| Baseline | 37.1 | 36.7 | 36.2 |
| AGVM | - | 37.1 (config) | 36.8 (config) |
The seg mAP of Mask R-CNN:
| Batch Size | 32 | 256 | 512 |
|---|---|---|---|
| Baseline | 34.8 | 34.4 | 33.9 |
| AGVM | - | 35.0 (config) | 34.6 (config) |
@article{xue2022large,
title = {Large-batch Optimization for Dense Visual Predictions},
author = {Zeyue Xue and Jianming Liang and Guanglu Song and Zhuofan Zong and Liang Chen and Yu Liu and Ping Luo},
year = {2022},
journal = {arXiv:2210.11078}
}