Add Support for multiple train loaders

CHANGELOG.md

@@ -9,6 +9,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
+- Add Support for multiple train loaders ([#1959](https://github.com/PyTorchLightning/pytorch-lightning/pull/1959))
+
 
 ### Changed
 

docs/source/multiple_loaders.rst

@@ -9,14 +9,16 @@ Multiple Datasets
 Lightning supports multiple dataloaders in a few ways.
 
 1. Create a dataloader that iterates multiple datasets under the hood.
-2. In the validation and test loop you also have the option to return multiple dataloaders
+2. In the training loop you can pass multiple loaders as a dict or list/tuple and lightning 
+   will automatically combine the batches from different loaders.
+3. In the validation and test loop you also have the option to return multiple dataloaders
    which lightning will call sequentially.
 
 ----------
 
 Multiple training dataloaders
 -----------------------------
-For training, the best way to use multiple dataloaders is to create a ``DataLoader`` class
+For training, the usual way to use multiple dataloaders is to create a ``DataLoader`` class
 which wraps your multiple dataloaders (this of course also works for testing and validation
 dataloaders).
 
@@ -59,6 +61,31 @@ dataloaders).
             # SAME
             ...
 
+However, with lightning you can also return multiple loaders and lightning will take care of batch combination.
+
+For more details please have a look at :attr:`~pytorch_lightning.trainer.trainer.Trainer.multiple_trainloader_mode`
+
+.. testcode::
+
+    class LitModel(LightningModule):
+
+        def train_dataloader(self):
+
+            loader_a = torch.utils.data.DataLoader(range(6), batch_size=4)
+            loader_b = torch.utils.data.DataLoader(range(15), batch_size=5)
+
+            # pass loaders as a dict. This will create batches like this:
+            # {'a': batch from loader_a, 'b': batch from loader_b}
+            loaders = {'a': loader_a,
+                       'b': loader_b}
+
+            # OR: 
+            # pass loaders as sequence. This will create batches like this:
+            # [batch from loader_a, batch from loader_b]
+            loaders = [loader_a, loader_b]
+
+            return loaders
+
 ----------
 
 Test/Val dataloaders

pytorch_lightning/trainer/data_loading.py

@@ -29,6 +29,9 @@
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.model_utils import is_overridden
 
+from pytorch_lightning.utilities.apply_func import apply_to_collection
+from pytorch_lightning.trainer.supporters import CombinedLoader
+
 
 class TrainerDataLoadingMixin(ABC):
 
@@ -137,6 +140,7 @@ def reset_train_dataloader(self, model: LightningModule) -> None:
             model: The current `LightningModule`
         """
         self.train_dataloader = self.request_dataloader(model.train_dataloader)
+
         if (self.overfit_batches > 0):
             if hasattr(self.train_dataloader, 'sampler') and isinstance(self.train_dataloader.sampler, RandomSampler):
                 rank_zero_warn('You requested to overfit but enabled training dataloader shuffling.'
@@ -147,13 +151,17 @@ def reset_train_dataloader(self, model: LightningModule) -> None:
         # debugging
         self.dev_debugger.track_load_dataloader_call('train_dataloader', dataloaders=[self.train_dataloader])
 
-        self.num_training_batches = 0
-
         # automatically add samplers
-        self.train_dataloader = self.auto_add_sampler(self.train_dataloader, shuffle=True)
+        self.train_dataloader = apply_to_collection(
+            self.train_dataloader, DataLoader, self.auto_add_sampler, shuffle=True)
+
+        # check the workers recursively
+        apply_to_collection(self.train_dataloader, DataLoader, self._worker_check, 'train dataloader')
+
+        # wrap the sequence of train loaders to a CombinedLoader object for computing the num_training_batches
+        self.train_dataloader = CombinedLoader(self.train_dataloader, self._multiple_trainloader_mode)
 
         self.num_training_batches = len(self.train_dataloader) if has_len(self.train_dataloader) else float('inf')
-        self._worker_check(self.train_dataloader, 'train dataloader')
 
         if isinstance(self.limit_train_batches, int) or self.limit_train_batches == 0.0:
             self.num_training_batches = min(self.num_training_batches, int(self.limit_train_batches))