Accepting serarete hyperparameters for testing phase

[dasayan05]

I have one script train.py for training + validation ..

class TrainPLModule(pl.LightningModule):
    def __init__(self, all, my, hps):
        self.save_hyperparameters()
        # all my hyper params can be accessed from self.hparams
    def training_step(self, ...): ...
    def validation_step(self, ...): ...

system = TrainPLModule(all=ALL, my=MY, ...)
trainer.fit(system)

.. another test.py for testing

class TestPLModule(TrainPLModule):
    def test_step(self):
        # self.hparams are all indeed accessible, BUT ..
        # I want new hparams specific to testing process
        self.hp.test_specific_hp # NOT AVAILABLE

system = TestPLModule.load_from_checkpoint('/trained/model',
                                            all=NEW_ALL, # overwrites old hparam, I know !
                                            # but this WON'T show up in self.hp in test_step()
                                            test_specific_hp=12
trainer.test(system)

Can I have some test specific HPs to show up in the test script ?

I know I can define all HPs in train file and overwrite them in test file, but my testing process is large and have lots of testing specific HPs (totally unnecessary for training) .. so I don't want to clutter my training script.

[tchaton]

Dear @dasayan05, interesting use-case :)

Mind opening a feature request, so we can brainstorm on how to properly support this.

Best,
T.C

[dasayan05]

Just did #8631