Closed wishforgood closed 1 year ago
What is the correct way for training a dino model?
I tried
python tools/train_net.py --config-file projects/dino/configs/dino_r50_4scale_12ep.py
like instructed in the doc, but got an error saysTypeError: get_default_optimizer_params() got an unexpected keyword argument 'lr_factor_func'
. I checked the code and found thatdino_r50_4scale_12ep.py
is really usingget_default_optimizer_params()
, which has not such an input.
It seems like your detectron2 is not the correct version, you should install d2 follow our official installation guide: https://detrex.readthedocs.io/en/latest/tutorials/Installation.html
which we keep the version of detectron2 fixed using github submodule @wishforgood
Sorry, I'm not able to install d2 following your guide, after installing that way, I run random codes and got an error like this
It seems like that the c++ part failed to be compiled.
So I tried to install a fixed pre-compiled version of d2, which I thought was the newest one: https://github.com/facebookresearch/detectron2/releases/tag/v0.6
Sorry, I'm not able to install d2 following your guide, after installing that way, I run random codes and got an error like this
It seems like that the c++ part failed to be compiled.
So I tried to install a fixed pre-compiled version of d2, which I thought was the newest one:
https://github.com/facebookresearch/detectron2/releases/tag/v0.6
This is not the newest one, it's only the newest package~, maybe you should figure out how to install their latest code, our guide is one way
Sorry, I'm not able to install d2 following your guide, after installing that way, I run random codes and got an error like this It seems like that the c++ part failed to be compiled. So I tried to install a fixed pre-compiled version of d2, which I thought was the newest one:
https://github.com/facebookresearch/detectron2/releases/tag/v0.6
This is not the newest one, it's only the newest package~, maybe you should figure out how to install their latest code, our guide is one way
OK, I understand. I will try again to compile it someway.
Sorry, I'm not able to install d2 following your guide, after installing that way, I run random codes and got an error like this It seems like that the c++ part failed to be compiled. So I tried to install a fixed pre-compiled version of d2, which I thought was the newest one:
https://github.com/facebookresearch/detectron2/releases/tag/v0.6
This is not the newest one, it's only the newest package~, maybe you should figure out how to install their latest code, our guide is one way
OK, I understand. I will try again to compile it someway.
the other way is that you can try to find the specific pull request and add the code into your own local package~
No need to recompile? OK, that's a good news, thanks. I will try it.
No need to recompile? OK, that's a good news, thanks. I will try it.
Yes, it's just pytorch code, find the specific pr about adding lr_factor_func
, and just add these code into your own local detectron2 package~
What is the correct way for training a dino model?
I tried
python tools/train_net.py --config-file projects/dino/configs/dino_r50_4scale_12ep.py
like instructed in the doc, but got an error saysTypeError: get_default_optimizer_params() got an unexpected keyword argument 'lr_factor_func'
. I checked the code and found thatdino_r50_4scale_12ep.py
is really usingget_default_optimizer_params()
, which has not such an input.