-
Notifications
You must be signed in to change notification settings - Fork 161
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Fast_SCNN based on the original config yaml cannot reproduce the mIoU? #18
Comments
so your batch size is 12 * 8? I only use batch 12 with 1 gpus which is consistent with paper. |
The BATCH_SIZE in the config file refers to the total batch_size or the batch_size on each gpu? My understanding is that this parameter is the batch_size on each gpu |
you are right, it is the batch_size on each gpu, i updated my reply. so only need 1 gpus to make total batch size be 12. |
Thank you for your reply. Are all the parameters in the configs file trained by 1 gpu by default? |
Only one GPU training and bs=12 is so slow, If I train with 4 GPUs and thus the batchsize=48? Any else should I change to reproduce 68.9 by training with 4gpus and bs=48 ? |
you can set batch_size=3 in config file, and use 4gpu to keep total batch size=12 |
no, i only trained 3 or 4 models. i think keep total batch size=12~16 will be ok. |
I mean bs=12 training is so slow, I just want train with larger batchsize . If I use bs=48 (12*4) and change lr=0.045x4 , can I reproduce 68.9? |
i have not tried this hyper parameters. |
hi, I see you get the trained model, dou you know where can i obtain the trained cityscapes_deeplabv3_plus_mobilenet model? |
https://github.com/LikeLy-Journey/SegmenTron#real-time-models |
Without any change, I train with 8 gpus but only get 59.0 mIoU on cityscapes. But using your trained fast_scnn_segmentron.pth can reproduce 68.9
Did you change something later about the config yaml?
Thanks for your nice work!
The text was updated successfully, but these errors were encountered: