I would suggest you start using cfg files and just have a debug.cfg etc. Right now it seems you are changing default values in your code all the time. This is very dangerous as you may introduce errors or forget to change things back. What I would suggest is:

[lucamarini22]

          I would suggest you start using cfg files and just have a debug.cfg etc. Right now it seems you are changing default values in your code all the time. This is very dangerous as you may introduce errors or forget to change things back. What I would suggest is:

1. Have toml cfg files. One is default which you never touch ideally and you can write a small test that checks that the code still works with it (what I do for this is create a mock dataset of a few images to train for one iteration just to sanity check I didn't break my training pipeline. Aside from that have a "debug.toml." or whatever you want for testing. not all have to be git-tracked either.
2. when you run a training run, with your outputs , logs etc copy the cfg file then you can always replicate what you did
3. use cfg to define as many of your hard-coded strings and constants as possible. (I see some hard-coded paths everywhere)

Originally posted by @gomezzz in #23 (comment)