search

YudeWang / SEAM

Self-supervised Equivariant Attention Mechanism for Weakly Supervised Semantic Segmentation, CVPR 2020 (Oral)
MIT License
539 stars 97 forks source link

Training the segmentation code #11

Closed halbielee closed 3 years ago

halbielee commented 4 years ago

Hello!

Thank you for sharing the excellent code.

I am trying to reproduce the performance you reported and I tried to train the result of the affinity network [Ahn et al.] with the segmentation code of https://github.com/itijyou/ademxapp

But I failed to train. Can you share the hyper-parameters or any change when you train? From the affinity net I found that he changed SGD to Adam with his work.

You may not remember, I need a little clue.

Thank you.

YudeWang commented 4 years ago

Hi @halbielee ,

The original segmentaion code is not available for me at present because the server is down and not fixed yet. I have a copy of the project but the hyper-parameters have been changed few times and I am not very clear about the exactly setting for the paper experiments. I will list the present parameters in that configuration file.

The segmentation model is Deeplabv1 with resnet38 backbone (maybe you need to add some Deeplab module on resnet38 from the link you given). Noting that the additional layers' learning rate is 10x than backbone layers. SGD with lr=0.001, weight_decay=0.0005, poly policy with 0.9, GPUS=4, batchsize=10, itr=20k, random crop_size=321 (maybe 448), random rescale [0.5-2.0], random flip, random color jitter multi-scale test [0.5, 0.75, 1.0, 1.25, 1.5, 1.75]

halbielee commented 4 years ago

Thank you @YudeWang

I will try that and let you know the result.!

TyroneLi commented 3 years ago

Thank you @YudeWang

I will try that and let you know the result.!

Hi~have you solved this problem? Which deeplab code repro you chosed finally? could you share a link or details with me? Thansk

halbielee commented 3 years ago

Hello @TyroneLi I could not solve this. When I modify this with DeepLab module, it shows 3-5 % lower performance than the normal WideResNet38 network.

YudeWang commented 3 years ago

@TyroneLi @halbielee Here is a screenshot of deeplabv1 file, maybe it will help you. image

Gzn520 commented 3 years ago

请问，在推断时，cam 和crf 是训练时得到的吗 为什么 没有看到生成呀？

Gzn520 commented 3 years ago

如何让cam可视化？

zhudahui commented 3 years ago

@Gzn520 请问您解决了吗

zbf1991 commented 3 years ago

@YudeWang I try to re-implement deeplab-resnet38 following your screenshot, but the performance is much lower, could you please share your segmentation code?

YudeWang commented 3 years ago

@TyroneLi @halbielee @zbf1991 @zhudahui Retrain step code is released here with approximate performance comparing to paper.

YudeWang commented 3 years ago

Any question about retrain step can be given in new repository and I will close this issue.