cityscapes model

Question

cityscapes model

vobecant opened this issue 3 years ago · comments

vobecant commented 3 years ago

Dear authors,

can you please share the weights of the model trained on the Cityscapes dataset?

Thank you very much in advance.

Best,
Antonin.

rstrudel · Answer 1 · Tue Sep 14 2021 17:14:00 GMT+0800 (China Standard Time)

Hi @vobecant ,
We uploaded the weights of the model on Cityscapes, please check it out!
Robin

vobecant · Answer 2 · Tue Sep 14 2021 17:21:27 GMT+0800 (China Standard Time)

Thanks @rstrudel !

vobecant · Answer 3 · Wed Sep 22 2021 16:28:07 GMT+0800 (China Standard Time)

Dear @rstrudel ,
I see that the model is Seg-L-Mask/16. As I recall, in one of your earlier paper versions, you said that you were not able to train larger models than Seg-B†-Mask/16. Can I please ask you how did you manage to train the Large variant? I tried to fit something to V100 but I was not able to train the Large model.
Also, I see in the variant.yml that you don't even use amp.
Thank you very much in advance!

rstrudel · Answer 4 · Wed Sep 22 2021 20:28:43 GMT+0800 (China Standard Time)

Hi @vobecant,
Sure, as specified in the paper for Cityscapes we train the mask transformer with 1 layer instead of two and it does fit memory on 8 V100 GPUs with 32Go of memory. We indeed added amp in the settings to speed up training. Unfortunately it made training really unstable and prone to NaN values after a while in some cases. Thus we just decided to turn it off and opt for the slower but more reliable full numerical precision.