training_epoch needs to return a "loss" key in the dict

📚 Documentation

Hi everyone!

In the docs detailing the usage of the logging functions training_epoch_end, if "loss": loss is not explicitly passed as a return value, then the code will fail.

The docs at https://pytorch-lightning.readthedocs.io/en/latest/experiment_reporting.html#log-metrics is not correct,

def training_epoch_end(self, outputs):
   loss = some_loss()
   ...

   logs = {'train_loss': loss}
   results = {'log': logs}
   return results

shall be changed to

def training_epoch_end(self, outputs):
   loss = some_loss()
   ...

   logs = {'train_loss': loss}
   results = {'loss': loss, 'log': logs} # <------------------------ Here is the change
   return results

Use case: I want to monitor the training metrics to check whether my network is able to overfit the data (for this functionality, refer to #1076 )

[rpatrik96]

Oh, you are right, I was using the old version; interestingly, I got no warning even on a server using the newest lightning version. Is it any way to enable explicit warnings about to be deprecated features. I like to maintain my code to be possibly up to date - although, it does not happen all the time, as in this case. Anyway, I am closing this issue.

[awaelchli]

training_epoch_end does not exist yet, right (#1076)? When you say the code fails, you probably used the old training_end which was on batches and now is renamed to training_step_end. There the loss key is required.

I don’t see why the loss key should be mandatory in training_epoch_end. As far as I know, validation_epoch_end also does not require a loss key in return dict.