How to return a final val loss in trainer?

[Data-drone]

What is your question?

Most optimisation packages ie Ray Tune / Hyperopt return the train loop to return a final accuracy for the optimiser to decide what to try next.

How do I do this with the Trainer module for Pytorch Lightning?

What's your environment?

• OS: Linux
• Packaging pip
• Version 0.7.6

[rajarajanvakil]

I too have the same issue , I want to return "val_loss" in validation step and "avg_val�_loss" from validation_epoch_end

[rajarajanvakil]

#321

`def validation_end(self, outputs):
avg_loss = torch.stack([x['batch_val_loss'] for x in outputs]).mean()
avg_acc = torch.stack([x['batch_val_acc'] for x in outputs]).mean()

    return {
      'val_loss': avg_loss,
      'val_acc': avg_acc, 
      'progress_bar':{'val_loss': avg_loss, 'val_acc': avg_acc }}`

[Data-drone]

I managed to achieve using callbacks and loggers but it doesn't work for ddp backend if doing distributed training. Think I need to include a manual gather step in there for that

[williamFalcon]

let’s formally support this somehow so it doesn’t have to be hacker around :)