trainer.fit() hangs if first defining a nn.Module and then passing it to pl.LightningModule

[zhiqiangdon]

This tutorial suggests using pl.LightningModule to define a system, i.e., first defining a nn.Module model and then passing it to pl.LightningModule like this:

class LitModel(LightningModule):
    def __init__(self, model: nn.Module = None):
        super().__init__()
        self.model = model

However, for the above case, I find trainer.fit() hangs after calling the configure_optimizers. Nothing is shown in the terminal and GPUs are unused.

If I change the definition as follows:

class LitModel(LightningModule):
    def __init__(self, config):
        super().__init__()
        self.model = MyModel(config)

The trainer.fit() works this time.

I am confused why the first definition causes trainer.fit() stuck, even the validation sanity check is not run.

[rohitgr7]

it should work fine. Mind sharing a reproducible example?

[zhiqiangdon]

Thanks @rohitgr7 ,

I have found the reason of hanging. It is due to that I called self.save_hyperparameters() without excluding the model. After I do self.save_hyperparameters(ignore=["model"]), the trainer.fit() works.