Improve docs for multiple dataloaders

[NielsRogge]

📚 Improve documentation

It would be great if you could extend this page by also explaining how to define the training_step and validation_step when using multiple dataloaders.

Currently, it's not that clear to me how to modify this. For example, when the train_dataloader is defined as follows:

class LitModel(LightningModule):

    def train_dataloader(self):

        loader_a = torch.utils.data.DataLoader(range(8), batch_size=4)
        loader_b = torch.utils.data.DataLoader(range(16), batch_size=4)
        loader_c = torch.utils.data.DataLoader(range(32), batch_size=4)
        loader_c = torch.utils.data.DataLoader(range(64), batch_size=4)

        # pass loaders as a nested dict. This will create batches like this:
        # {'loader_a_b': {'a': batch from loader a, 'b': batch from loader b},
        #  'loader_c_d': {'c': batch from loader c, 'd': batch from loader d}}
        loaders = {'loaders_a_b': {'a': loader_a, 'b': loader_b},
                   'loaders_c_d': {'c': loader_c, 'd': loader_d}}
        return loaders

, does the API of training_step still look like training_step(self, batch, batch_idx)? Or how can I access the multiple batches?

Thanks!

[tchaton]

Dear @NielsRogge,

Yes, the training_step doesn't change.

class LitModel(LightningModule):

    def train_dataloader(self):

        loader_a = torch.utils.data.DataLoader(range(8), batch_size=4)
        loader_b = torch.utils.data.DataLoader(range(16), batch_size=4)

        return {'a': loader_a, 'b': loader_b}

   def training_step(self, batch, batch_idx):
        batch_a = batch["a"]
        batch_b = batch["b"]
       ....         

Best,
T.C

[NielsRogge]

Thanks for improving this!