search

tztztztztz / eql.detectron2

The official implementation of Equalization Loss for Long-Tailed Object Recognition (CVPR 2020) based on Detectron2. https://arxiv.org/abs/2003.05176
Apache License 2.0
202 stars 17 forks source link

How to reproduce the results in your paper? #8

Closed z-x-yang closed 4 years ago

z-x-yang commented 4 years ago

I want to reproduce your paper's results, but it seems like you changed some settings (such as Scale jitter) in this repository.

May I ask for the code with the original setting in your paper? Or how to modify this repository to reach that?

Thanks in advance!

tztztztztz commented 4 years ago

Sorry, the origin code is based on a internal codebase, which is confidential.

If you want to match the origin setting in this repo, here are some modifications to make:

  1. single scale training INPUT: MIN_SIZE_TRAIN: (800, )

  2. 2000 training proposals RPN: POST_NMS_TOPK_TRAIN: 2000

  3. class agnostic mask head ROI_MASK_HEAD: CLS_AGNOSTIC_MASK = True

  4. Use smooth L1 loss instead of L1 RPN: SMOOTH_L1_BETA = 1.0 / 9.0 ROI_BOX_HEAD: SMOOTH_L1_BETA = 1.0

z-x-yang commented 4 years ago

THX!