Getting predictions and gold labels from trainer.test()

[ZhaofengWu]

Is there a way to obtain predictions and gold labels from trainer.test()? I have access to these in test_epoch_end(), but since this function doesn't return, I can't pass them outside. Options that I've considered:

1. Use self.log("preds", preds), but apparently torch tensors are automatically reduced.
2. Use trainer.predict(), but since I'm already doing trainer.test(), having to run the forward pass on the evaluation data twice seems wasteful.
3. Set self._some_custom_property_{preds,labels} in test_epoch_end(), and access these properties outside. I'd go for this if there's no better alternative, but it sounds like a hack.