bug: logging to wandb does not work

[MartinBernstorff]

🤖 This is the current blocker for a full hyperparameter search.

Currently have to write the manuscript, so unable to debug in detail. We need a permanent solution shared across projects, therefore the issue here.

Is this related to the latest internet shutdown? Maybe so, but we already open the URLs I could find from wandb https://api.wandb.ai.

• Train single model - seemed to work fine
• Run full model training using the train_models_in_parallel script
• It only runs exactly 10 models, one for each lookahead/model architecture combination
  • I.e. it doesn't start more than one model for each combination, even though we get
    • This implies there's something wrong with our hyperparameter search.
      • Running the command directly from the command line
        • Seems that this does not sync correctly. Stuck on the last line.
    • Suggests that it's a wandb/hydra interaction that's causing the problem, but only when using --multirun
      • Testing without multirun
        • Huh, same problem!
    • Testing again, but just a simple "train model", i.e. circumventing Hydra's CLI interface
      • Still not working, that's weird!

Appears we might be hitting rate limiting, since training a single model worked fine the first time, and then didn't work?

Proposed next steps for debugging:

• Check if we can train and upload even a single model (note that syncing continuous after

• If we can, but cannot train in parallel, appears it's a wandb problem. Potential next steps:

  • Write performance to disk and drop wandb support
  • Switch to another provider (local/remote)

If anyone is up for debugging, they're more than welcome to go ahead and collect thoughts here! @HLasse, @sarakolding, @signekb, @bokajgd, @erikperfalk.

[MartinBernstorff]

Har skreve til wandb support for at høre, om der er andre hostnames vi skal åbne for.