megvii-basedetection/autoassign

Pytorch implementation of "AutoAssign: Differentiable Label Assignment for Dense Object Detection"

AutoAssign: Differentiable Label Assignment for Dense Object Detection

This is a PyTorch implementation of the AutoAssign paper:

@article{zhu2020autoassign,
  title={AutoAssign: Differentiable Label Assignment for Dense Object Detection},
  author={Zhu, Benjin and Wang, Jianfeng and Jiang, Zhengkai and Zong, Fuhang and Liu, Songtao and Li, Zeming and Sun, Jian},
  journal={arXiv preprint arXiv:2007.03496},
  year={2020}
}

Get Started

1. install cvpods following the instructions

# Install cvpods
git clone https://github.com/Megvii-BaseDetection/cvpods
cd cvpods 
## build cvpods (requires GPU)
pip install -r requirements.txt
python setup.py build develop
## preprare data path
mkdir datasets
ln -s /path/to/your/coco/dataset datasets/coco

2. run the project

cd auto_assign.res50.fpn.coco.800size.1x

# train
pods_train --num-gpus 8

# test
pods_test --num-gpus 8
# test with provided weights
pods_test --num-gpus 8 MODEL.WEIGHTS /path/to/your/model.pth

Results

Model	Multi-scale training	Multi-scale testing	Testing time / im	AP (minival)	Link
AutoAssign_Res50_FPN_1x	No	No	53ms	40.5	download

autoassign's Issues

Warning and different result

I downloaded the pretrained model provided. During inference I'm getting the following warning. It results worse AP ~29. Also, could you upload the other models with different backbones that you share on the paper?

WARNING [02/12 13:37:59 c2.checkpoint.checkpoint]: 'backbone.top_block.p6.weight' has shape (256, 2048, 3, 3) in the checkpoint but (256, 256, 3, 3) in the model! Skipped.
[02/12 13:37:59 c2.checkpoint.checkpoint]: Some model parameters are not in the checkpoint:
backbone.top_block.p6.weight

Training is slow.

I used AutoAssign to train my data set, but the speed of training was too slow.Is that normal?

Performance on VOC

Hi, 我最近在研究label assign，实验发现使用默认设置下的AutoAssign在VOC上的表现会低于GFL 3个点以上的AP。请问你们有在VOC上尝试过吗？我使用的是mmdet官方repo中的AutoAssign，最大epoch为4，学习率下降在第3 epoch之后，这是mmdet VOC的标准设置。GFL可取得51.8的AP，但AutoAssign只能取得48.4。请问有什么超参会对性能影响较大？

how can I debug code ？

Hi, thanks for your nice work.

I'm extremely interesting in the ImpObj branch in your paper, but you code is based on cvpods base,
so how could I debug code ？

Thanks ！

loss norm

Thank you for your excellent work. I have a question, what does "loss_norm" mean, this seems not mentioned in the original paper

I get mAP40.2, is this result normal?

Thank you for your outstanding work!
I recently ran this code and got a result of mAP40.2, Is this the normal fluctuation range ?

Training on VOC

Hello, thanks for your great work!

Can you provide a config.py file for the training on PASCAL VOC dataset.

Thanks