Implement `burn-report` for Training Progress Reporting to CometML and Other Dashboards

[antimora]

Description of the Feature

The burn-report is intended to comprise a suite of dashboard API clients, beginning with an implementation for CometML.

Motivation for the Feature

There are several available clients for reporting training progress in Python, such as CometML and TensorBoard. Nevertheless, a reliable counterpart for Rust is currently absent. The burn-report feature is being proposed to address this gap and offer a versatile tool for Rust users.

Proposed Solution

A three-step solution is suggested to implement this feature:

1. Develop necessary abstractions for the presentation of training data.
2. Construct async clients for CometML, using their Rest APIs to facilitate interaction.
3. Build an async client capable of storing reports locally to improve data accessibility and redundancy.

[wdoppenberg]

I'd say a good place to get inspired by would be PyTorch Lightning's logger module: https://github.com/Lightning-AI/lightning/tree/master/src/lightning/pytorch/loggers

[antimora]

I'd say a good place to get inspired by would be PyTorch Lightning's logger module: https://github.com/Lightning-AI/lightning/tree/master/src/lightning/pytorch/loggers

Thank you for link. I will check it out!

[wangjiawen2013]

Good feature!
Where can we find the usage and examples of this feature ? We'd like to get the visualization like tensorboard with burn:

[laggui]

@wangjiawen2013 this is not currently available. It's a feature request, though it hasn't seen a lot of activity.

Regarding tensorboard specifically, they use tfevents which iirc contains event protobuf mesasges that tensorboard can read. Don't think this is the most compatible format for a general visualization.