search

Arthur-Null / SRD

Official pytorch implementation of Spatial Relation Decomposition method (AAAI 23)
20 stars 2 forks source link

about utild #1

Open dituihuagui opened 1 year ago

dituihuagui commented 1 year ago

Dear author, the training code of SRD is written by utild, but I have found some problems. For example, some optional hyper-parameter can't pass the type check. Do you have any version of SRD's code which didn't use utild? Thanks

Arthur-Null commented 1 year ago

Can you explain in detail what error you meet? Currently, we do not have a version that does not use utilsd. However, You can easily implement a custom entry without using utilsd.

dituihuagui commented 1 year ago

“utilsd.config.exception.ValidationError: RuntimeConfig: None failed to pass type check of typing.Optional[pathlib.Path]” This is a problem of the hyper-parameter named "output_dir, checkpoint_dir,tb_log_dir" which I have encountered.
Another problem is the hyper-parameter named early_stop. It can't pass the type check because utilsd requires it's type to be int, but it's real type is Optional[int]. Do you build the utilsd by modifing the NNI's code? And if I want to rewrite the training code, should I read the source code of utilsd or the NNI? Thank you very much.

Arthur-Null commented 1 year ago

I think building utilsd from source would fix your problem. This repo does not depend heavily on utilsd as I only use it to initialize the network and models, etc. You can actually easily write your own entrance without using utilsd or NNI. You only need to write your own main file to replace forecaster/entry/tsforecast.py.