Where is the loggers.TrainsLogger?

[igormq]

Hey folks, you removed the loggers from v.0.2.0 to v0.2.1 and migrate all loggers to the core. However, I couldn't find the TrainsLogger. What did happen?

Thanks!

[bmartinn]

@Borda could you help? I could not locate it either...

[Borda]

@Borda could you help? I could not locate it either...

@williamFalcon ^^

[NumesSanguis]

Just ran into this issue. Upgraded my code for PyTorch Lightning 1.0.0rc4 and also updated Bolts to match it. Suddenly my code doesn't run because the Trains Logger is gone. I traced the relevant commit to this one: 91393ea

I hope that Lightining (Bolts) have plans for a more robust review processes, involving multiple people, before things are removed from master? As this deletion has been made without any (public) discussion: #215

Or least clearer communication to the relevant 3rd parties, because over at Trains they didn't know how this happened. And this already occurred before when the Trains logger was moved (due to a perceived issue with their code) from Lightning to Bolts, without giving them a chance to comment on it: #86

[williamFalcon]

The Trains logger was moved to bolts a long time ago. the code is still available there. We needed to stabilize 1.0.0 and trains was causing a lot of issues on our build.

Please open the issue there so we can take a look at it

[Ingwar]

@williamFalcon , sorry, I'm confused here. As far as I understand, this is an issue about disappearance of Trains logger from bolts, not from the main lightning. So, it's unavailable in the latest release of pytorch-lightning-bolts and it's unclear, was it a mistake or a deliberate decision. Could you, please, shed a bit of light on it ? :)

[edenlightning]

As communicated offline, to stabilize the repo we are migrating some integration to their 3rd party libraries. Hope we can resolve this soon!

[bmartinn]

@williamFalcon

Please open the issue there so we can take a look at it

Okay, I will PR the TrainsLogger in the main repo, back with the other loggers. It seems 'bolts' is now more of a recipe-book for SOTA models, which definitely is not the right place for loggers to be.

[williamFalcon]

@bmartinn, as @edenlightning mentioned, we are not looking to add loggers back into the main lightning repo because of the maintenance cost on our end. We think it would be great for you to add that integration on your end. @edenlightning can talk you through the details of what that looks like.

looking forward to making sure your users have a good experience!

Please email us at [email protected] or [email protected] to discuss further!
(sorry, this was locked by mistake)

[bmartinn]

@igormq, @Ingwar, @NumesSanguis, @S-aiueo32, @odedbd, Well guys I come bearing good news :D
@williamFalcon forgot to mention that starting PL v1.0 all logs are automatically logged by default to Tensorboard.
Now since train automagically connects with TB (as well as other frameworks) , it means there is no need for any specific logger whatsoever.

Just create a trains Task and you are good to go:

from trains import Task
...
if __name__ == '__main__':
    Task.init(project_name="examples", task_name="lightning checkpoint tensorboard and argparse")

For example, I took the mnist example from the PL repo, and added the Task.init line.

With that, any call to log(...) will be automatically logged by both the TB and the trains-server (see scalars, console)
trains also automatically logs the argparse so you have your full configuration logged as well (see configuration).
Obviously the full environment (code/git/packages) is also viewable (execution details)
Lastly, since all torch.save/load commands are logged, you can see all your model checkpoints (see models)
(as usual if you pass output_uri='s3://bucket/dir' to Task.init, these models will be automatically uploaded to a dedicated directory in the selected bucket, and of course this also applies for gs://, azure://, http:// or even just a shared folder)

Feel free to test with the latest RC, also verified that running remotely with trains-agent allows you to override any if the argparse arguments, as well as the output_uri.

pip install trains==0.16.3rc0

BTW:
We will be working on an even better Logger solution, more on that in the near future 🚆 😄

[NumesSanguis]

@bmartinn Thanks for the update! Does trains==0.16.3rc0 work with Trains Server v0.16.1 (we're using the Docker version)?

[bmartinn]

@NumesSanguis trains will support any combination of python package and trains-server , otherwise how could you reproduce old experiments :)
BTW: There is nothing special in version 0.16.3rc0 it is just the latest one, so I tested it with it. Any trains version should basically work the same way 😃