Context manager for common event metadata.

[tillahoffmann]

I often have code that looks like this.

writer.add_scalar("foo", some_value, global_step=step)
writer.add_scalar("bar", another_value, global_step=step)
writer.add_scalar("baz", yet_another_value, global_step=step)

There is some repetition with specifying the global_step, and I often forget to specify the step which leads to unexpected results in tensorboard (that's of course a user error). It might be possible to address both using a context manager like so.

with writer.use_global_step(step):
    writer.add_scalar("foo", some_value)
    writer.add_scalar("bar", another_value)
    writer.add_scalar("baz", yet_another_value)

Where global_step now defaults to whatever is specified as the argument to use_global_step (or a better function name). It would also be possible to incorporate the walltime parameter with a signature like use_metadata(*, global_step=None, walltime=None), where the user can specify global_step, walltime, or both.

Is this of interest? Happy to send a PR if so.

Thank you for the great library!

[lanpa]

@tillahoffmann Please sent a PR! I recently changed the package config of this repository to uv, hope that doesn't mess things up while developing new features.