I am trying to train SE-SSD on my custom dataset, however the training requires too much gpu memory.
In other issues you suggest using a pre-trained teacher network. (I tried using weights available on your repo but the model and loaded state dict do not match.)
Could you please elaborate on how to train only the teacher network to use it for the actual training after ?
Should I use the CIA-SSD repo ? Or use this repo and change a few things in the config file ?
Edit: I found how to do it (i.e. changing includes in a few init.py files and changing some things in the config file
I think a slightly more detailed readme.md would be welcome however :)
Hello,
Thanks for your work.
I am trying to train SE-SSD on my custom dataset, however the training requires too much gpu memory. In other issues you suggest using a pre-trained teacher network. (I tried using weights available on your repo but the model and loaded state dict do not match.)
Could you please elaborate on how to train only the teacher network to use it for the actual training after ? Should I use the CIA-SSD repo ? Or use this repo and change a few things in the config file ?
Edit: I found how to do it (i.e. changing includes in a few init.py files and changing some things in the config file I think a slightly more detailed readme.md would be welcome however :)