Sampler (#1318)

* sampler

* sampler

* sampler

* check for dataloader type

* check for dataloader type

Author: williamFalcon

CHANGELOG.md

@@ -24,6 +24,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
+- Does not interfere with a default sampler ([#1318](https://github.com/PyTorchLightning/pytorch-lightning/pull/1318))
 - Enhanced load_from_checkpoint to also forward params to the model ([#1307](https://github.com/PyTorchLightning/pytorch-lightning/pull/1307))
 - Made `evalaute` method private >> `Trainer._evaluate(...)`. ([#1260](https://github.com/PyTorchLightning/pytorch-lightning/pull/1260))
 

pytorch_lightning/trainer/data_loading.py

@@ -74,6 +74,15 @@ def _percent_range_check(self, name: str) -> None:
             raise ValueError(msg)
 
     def auto_add_sampler(self, dataloader: DataLoader, train: bool) -> DataLoader:
+
+        # don't do anything if it's not a dataloader
+        if not isinstance(dataloader, DataLoader):
+            return dataloader
+
+        # don't add sampler when user gives one
+        if dataloader.sampler is not None:
+            return dataloader
+
         if self.use_ddp or self.use_ddp2 or self.use_tpu:
             dl_args = {
                 'dataset': dataloader.dataset,

pytorch_lightning/trainer/distrib_parts.py

@@ -250,8 +250,11 @@
 .. note:: When running in DDP mode, any errors in your code will show up as an NCCL issue.
  Set the `NCCL_DEBUG=INFO` flag to see the ACTUAL error.
 
-Finally, make sure to add a distributed sampler to your dataset. The distributed sampler copies a
- portion of your dataset onto each GPU. (World_size = gpus_per_node * nb_nodes).
+Normally now you would need to add a distributed sampler to your dataset, however
+Lightning automates this for you. But if you still need to set a sampler Lightning will
+not interfere nor automate it.
+
+Here's an example of how to add your own sampler (again no need with Lightning).
 
 .. code-block:: python