search

feipanir / IntraDA

Unsupervised Intra-domain Adaptation for Semantic Segmentation through Self-Supervision (CVPR 2020 Oral)
https://arxiv.org/pdf/2004.07703.pdf
MIT License
270 stars 35 forks source link

low performance #5

Closed jianlong-yuan closed 3 years ago

jianlong-yuan commented 4 years ago

I follow your step python train.py --cfg ./configs/advent.yml python entropy.py --best_iter BEST_ID --normalize False --lambda1 0.67 python train.py --cfg ./intrada.yml

But i can not reproduce your performance. For the first stage, i got 41.98 mIoU For the second stage, i got 44.57 mIoU.

Could you help me?

feipanir commented 4 years ago

Hi are you doing the experiment from GTA5->Cityscapes? I guess the problem is from the first stage that you got 41.98, which is a bit low than what is expected. I suggest you to train more iterations in the first stage. In our experiments, we got BEST IOU at around 80k iteration or so. And also double check your learning rate, image size, pytorch version, etc.

jianlong-yuan commented 4 years ago

I am doing the experiment from GTA5->Cityscapes. I just use your default settings. How much miou do you achieve for the first stage?

image

image image image

By the way, pytorch version is 1.5, CUDA10.1

feipanir commented 4 years ago

We conduct the experiment on pytorch=0.4.1 and our best miou would be around 43.8 for the inter-da stage. Right now you are using pytoch=1.5. I guess it is probably due to pytorch version but I am not really sure. I will double check it on my local server.

jianlong-yuan commented 4 years ago

With pytorch0.4.1, i got 42.4miou at 5w iters. I have finished 10w iters.

feipanir commented 4 years ago

It's still a bit low than expected. We usually got 43.8± 0.2. Actually our experiments setup of inter-da keep the same as ADVENT.

feipanir commented 3 years ago

A little concern is also from the CUDA version. As I tried on my machine, by using CUDA=9.2 I got the same performance, rather it becomes lower when I switch to CUDA=10.2. So be careful about which cuda version is used currently.

xdeng7 commented 3 years ago

With pytorch0.4.1, i got 42.4miou at 5w iters. I have finished 10w iters.

Hi, I was using pytorch1.5.1 and CUDA10.2. I was able to achieve 43.75 for ADVENT, the first stage. And 46.28 for the second stage adaptation.

feipanir commented 3 years ago

Hi @xdeng7 thank you for your effort and sharing this information! Regards