Experiment detail

[2448845600]

Hi, this project is very helpful. But I would like to know some details:

1. the backbone you provided named resnet50_v2,pth and resnet101_v2.pth. Did you train this backbone from init for few-shot segmentation task?
2. I get same result with pre-trained model. But the model training by myself is lower about 2% mIoU. Is the config same with the paper setting?

Thanks.

[tianzhuotao]

@2448845600

The two pre-trained backbones are directly provided by the author of PSPNet . They were trained on ImageNet, and we did not tune them specifically for the few-shot segmentation task. You can find the same checkpoints at this cite.

For the reproduction issue, we have found that different experimental environments cause performance variance. The reproduced details/logs/checkpoints can be found in this issue.

With the same python dependencies/environments and the provided config files, this repo can achieve better performance than that reported in our paper.

Be free to contact me if you have further questions. Thank you~

[2448845600]

Hi, I reproduce the results now, thanks! The results is close to your paper when I use single-gpu (1*3090), However, if I use multi-gpu (4*3090), the mIoU is lower about 2%.
Could you please give me some ideas about multi-gpu?

[tianzhuotao]

@2448845600

We did not implement the multi-GPU training, because we did not include BN layers in some modules and therefore we do not know whether the sync-BN layers will adversely affect the performance.

BN layers should be replaced with the sync-BN layers if you want to implement the multi-GPU training. You can take references from either the PyTorch tutorial or the official implementation of PSPNet that our code was built upon.

[2448845600]

Thanks!