How often is train_dataloader() called for a LightningDataModule?

[EricWiener]

Hi,

I'm seeing my model overfit and I'm trying to figure out what's going wrong. I believe that the model isn't being trained on the entire train dataset. I define my LightningDataModule as follows:

class MyDataModule(pl.LightningDataModule):
    def train_dataloader(self):
        dataset = MyDataset()
        return DataLoader(MyDataset, num_workers=4)

trainer = Trainer(...)
model = Model(...)
data_module = MyDataModule(...)
trainer.fit(model, data_module)

Is it okay that I am re-initiating my dataset every time train_dataloader is called? I printed out when train_dataloader is called and it seems to only be called once and not get reset when switching from train -> val. Could I confirm this is the actual behavior and there isn't any possibility that MyDataset() will end up being re-initialized/reset somehow for training data?

Thank you!

[rohitgr7]

yes it will be called only once unless you set Trainer(reload_datalaoders_every_n_epochs=int > 0).