Some general clone clean up that should be addressed:
[x] Better location for wandb dir
[x] The warning logger.warning("Length of on-the-fly transformed dataset is not really fixed.") is popping up too much. Try to decide if this is strictly necessary.
[x] The complete run config (i.e. result of ... --cfg job --resolve) is not being logged on wandb. It should be.
Problem seems to be more of a wandb bug. When settings subconfigs (e.g. optimizer=class-recon) for single runs, it will correctly fill out the dictionary on wandb.ai, but somehow subconfig dicts are getting overwritten with key names during sweeps. Seems like a bug that may resolve itself.
[x] Torchinfo output should be dumped to a file on initialization.
[x] Fix command=clean not really detecting the existence of the dir properly
The problem is actually that even scan creates a directory, which isn't great but I'm not sure is worth avoiding.
Some general clone clean up that should be addressed:
logger.warning("Length of on-the-fly transformed dataset is not really fixed.")is popping up too much. Try to decide if this is strictly necessary.The complete run config (i.e. result of... --cfg job --resolve) is not being logged on wandb. It should be.optimizer=class-recon) for single runs, it will correctly fill out the dictionary onwandb.ai, but somehow subconfig dicts are getting overwritten with key names during sweeps. Seems like a bug that may resolve itself.Fixcommand=cleannot really detecting the existence of the dir properlyscancreates a directory, which isn't great but I'm not sure is worth avoiding.