Skip to content

Weights and Biases instead of Tensorboard (PR request) #5733

Description

@alexge233

Is your feature request related to a problem? Please describe.
I'm working as part of a team that runs multiple ml-agents on our own PCs and also plan on placing them on VPS. We want to collate and compare results using wandb instead of tensorboard as this will allow us to work better as a team.

Describe the solution you'd like
I've implemented a WandbWriter in my local git clone under ml-agents/mlagents/trainers/stats.py. I've added a simple unit test in ml-agents/mlagents/trainers/tests/test_stats.py, and injected the writer in ml-agents/mlagents/plugins/stats_writer.py so it becomes available. I haven't actually tested it with the samples/examples, only locally.

Describe alternatives you've considered
I can imagine the alternative to be to package my version of mlagents and distribute it to my colleagues, but that's less than ideal. It also means I'll have to copy, merge and maintain an updated copy. I'd rather not do that.

Additional context
This repo seems to have reached its maximum forks, and therefore I could not Fork + PR. Hence opening this issue. Happy to have the changes reviews, rejected, commented on, etc. but I can't seem to do that as things stand.
Also, I may need some guidance on how exactly you expect the format for the kwargs being passed to the writer to look like.

Metadata

Metadata

Assignees

Labels

requestIssue contains a feature request.

Type

No type
No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions