The simplest way to do trainier.predict on an arbitrary file after each fit loop

[whaowhao]

I have been trying to do something extremely simple -- predict on an arbitrary file after each fit loop using pl 1.5.0 but failed. This is the custom FitLoop. I understand there are more complicated ways to achieve it using baal and flash, but is there a reason this simple feature is not supported?

class ActiveLearningFitLoop(FitLoop):
    def __init__(self, min_epochs = 1, max_epochs = 1000):
        super().__init__(min_epochs, max_epochs)

    def on_advance_end(self):
        super().on_advance_end()
        args = self.trainer.lightning_module.hparams
        dl_active_learn = DataLoader(
            torch.load(os.path.join(args.cache_dir, "ds_active_learn.pt")), 
            batch_size=args.batch_size, 
            shuffle=False, 
            pin_memory=True, 
            num_workers=args.num_workers
        )
        self.trainer.predict(dataloaders=dl_active_learn)