Training loss is not being logged for pytorch-lightning version greater 0.8.5

Hi, when using wandb with pytorch-lightning I experienced that the training loss is not automatically logged anymore. I tried the versions > 0.8.5. The test loss is still being fetched. Since this problem also affects other logging integrations, e.g. Neptune.ai (verified), I am not sure on which side the logging is failing.

Tried code:

from pytorch_lightning.loggers import WandbLogger
wandb.init(project="project")
trainer = pl.Trainer(max_epochs=5, logger=wandb_logger, gpus=0, weights_summary='full')

Happened to anyone else?

[borisdayma]

I’m not sure which version you use but the pattern for logging is now with self.log (see the docs)

[sneakyPad]

Hi @FraPochetti! Sorry @borisdayma for not getting back at this. I remember that I tried self.log but at that time I was logging a lot, and if I recall correctly, I was only able to log a single variable with it? (Not quite sure)

@FraPochetti I ended up with the following:

tensorboard_logs = {'train_loss': batch_loss} 

return {'loss': batch_loss, 'log': tensorboard_logs}

I think the loss in the return statement never got logged and I was just a bit careless. I think the trick was using tensorboard_logs for log. Let me know if this works.

You can also have a look at the full code here: github.com/sneakyPad/decoding-latent-space-rs/blob/master/models/movies_vae.py