How to train for a new dataset ?

VITA-Group / FasterSeg

[ICLR 2020] "FasterSeg: Searching for Faster Real-time Semantic Segmentation" by Wuyang Chen, Xinyu Gong, Xianming Liu, Qian Zhang, Yuan Li, Zhangyang Wang

MIT License

524 stars 107 forks source link

How to train for a new dataset ? #36

Closed MohamedAliRashad closed 3 years ago

MohamedAliRashad commented 4 years ago

So i have been playing a little bit with the config_search.py to adapt my new dataset to the pipeline but i kept getting this error

RuntimeError: merge_sort: failed on 2nd step: device-side assert triggered

After some research i knew it's a problem with the configuration i did to the dataset but the problem is that i don't know what every parameter do, so can anyone elaborate on this ?

chenwydj commented 4 years ago

Hi @MohamedAliRashad!

Thank you for your interest in our work!

Actually I have no idea what the problem is... The only place I found may be related to the merge_sort is the OHEM loss, but I'm not sure if this the where the error is raised.

MohamedAliRashad commented 3 years ago

@chenwydj Can you provide a description for every parameter in config.py ?

chenwydj commented 3 years ago

Hi @MohamedAliRashad! I have updated the comments for the configure files.