Trainer.profiler should allow str

[carmocca]

🚀 Feature

Currently, the profile argument of the Trainer is bool. It should be Union[bool, str].

Motivation

It would avoid having to set the profiler manually by creating and passing the object to the Trainer.
It is also the way auto_scale_batch_size and auto_lr_find work.

Pitch

# equivalent to --profiler true
trainer = Trainer(profiler=True)

# equivalent to --profiler simple
trainer = Trainer(profiler=SimpleProfiler())

# equivalent to --profiler advanced
trainer = Trainer(profiler=AdvancedProfiler())

[Borda]

@jeremyjordan any thoughts?

[jeremyjordan]

Sounds good to me, want to submit a PR @carmocca?

It should be Union[bool, str].

Technically, it should be Union[bool, str, BaseProfiler].

[carmocca]

Sounds good to me, want to submit a PR @carmocca?

Bit busy right now, maybe at a later point.

@Borda Tag it "good first issue" in the meantime?

[ddrevicky]

@Borda Assign this to me please :).

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!

[ananthsub]

if you're using Hydra, this instantiation via config becomes a lot easier. This has parallels to the distributed_backend / plugin interfaces too, which are increasingly supporting explicit instantiations passed to the trainer. Supporting string with lightning-provided implementations along with custom classes passed to the Trainer seems like a nice convergence! @williamFalcon