Saving a model checkpoint file with updated parameters

[sunburntfish]

What's the correct way to update a saved checkpoint file if I change the hyperparameters? For example if I have something like this:

class MyModel(LightningModule):
    def __init__(self, param1: int):
        self.param1 = param1
        self.save_hyperparameters()

model = MyModel()

callbacks = [ModelCheckpoint(monitor='val_loss')]
trainer = Trainer.from_argparse_args(args=args, logger=logger, callbacks=callbacks)

trainer.fit(model=model, datamodule=data_module)

I can then reload the checkpoint no problem with:

MyModel.load_from_checkpoint(checkpoint_file)

But if I change my constructor:

class MyModel(LightningModule):
    def __init__(self, param1: int, param2: str):
        self.param1 = param1
        self.param2 = param2
        self.save_hyperparameters()

the call to load from checkpoint will complain that it is missing the param2 from parameter list.

I can load it if I pass the parameter at load time:

MyModel.load_from_checkpoint(checkpoint_file, param2)

What I want at this point is a method that just saves the updated model with the new parameter to a new checkpoint file, e.g.:

MyModel.save_checkpoint(checkpoint_file)

But it seems the only way to save the checkpoint is through using the Trainer after the model has been internally coupled with the trainer. This is really awkward as I don't want to call trainer.fit(model) and if I use trainer.test(model) and save afterwards with trainer.save_checkpoint(checkpoint_file) the file is much smaller than my original checkpoint (I think because it only saves the weights?).

What's the right way to update the model checkpoint file if I change the model parameters?