data transfer model hook (+ refactor) (#1756)

* refactor and added hook


variant a


variant b


add test


revert rename


add changelog


docs

* resolve merge duplication

* overridden typo

* fix test

* tpu id

* raise if TPU not available

* re-use apply_to_collection function for parsing collections

* comment

* make utility function available to user

* documentation

* move changelog entry to top

* fix tpu transfer call

* fix call

* remove hardcoded string

* improve test

* call model hook by default

* Apply suggestions from code review

* rename utility function

Co-authored-by: Jirka Borovec <[email protected]>

Author: awaelchli

CHANGELOG.md

@@ -22,6 +22,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Speed up single-core TPU training by loading data using `ParallelLoader` ([#2033](https://github.com/PyTorchLightning/pytorch-lightning/pull/2033))
 
+- Added a model hook `transfer_batch_to_device` that enables moving custom data structures to the target device ([1756](https://github.com/PyTorchLightning/pytorch-lightning/pull/1756)).
+
 ### Changed
 
 - Allow user to select individual TPU core to train on ([#1729](https://github.com/PyTorchLightning/pytorch-lightning/pull/1729))

pytorch_lightning/__init__.py

@@ -60,8 +60,8 @@
         'Trainer',
         'LightningModule',
         'Callback',
-        'data_loader'
-        'seed_everything'
+        'data_loader',
+        'seed_everything',
     ]
 
     # necessary for regular bolts imports. Skip exception since bolts is not always installed

pytorch_lightning/core/hooks.py

@@ -3,6 +3,8 @@
 import torch
 from torch import Tensor
 from torch.optim.optimizer import Optimizer
+from pytorch_lightning.utilities import move_data_to_device
+
 
 try:
     from apex import amp
@@ -153,3 +155,48 @@ def backward(self, use_amp, loss, optimizer):
                     scaled_loss.backward()
         else:
             loss.backward()
+
+    def transfer_batch_to_device(self, batch: Any, device: torch.device) -> Any:
+        """
+        Override this hook if your :class:`~torch.utils.data.DataLoader` returns tensors
+        wrapped in a custom data structure.
+
+        The data types listed below (and any arbitrary nesting of them) are supported out of the box:
+
+        - :class:`torch.Tensor`
+        - :class:`list`
+        - :class:`dict`
+        - :class:`tuple`
+        - ``torchtext.data.Batch`` (COMING SOON)
+
+        For anything else, you need to define how the data is moved to the target device (CPU, GPU, TPU, ...).
+
+        Example::
+
+            def transfer_batch_to_device(self, batch, device)
+                if isinstance(batch, CustomBatch):
+                    # move all tensors in your custom data structure to the device
+                    batch.samples = batch.samples.to(device)
+                    batch.targets = batch.targets.to(device)
+                else:
+                    batch = super().transfer_batch_to_device(data, device)
+                return batch
+
+        Args:
+            batch: A batch of data that needs to be transferred to a new device.
+            device: The target device as defined in PyTorch.
+
+        Returns:
+            A reference to the data on the new device.
+
+        Note:
+            This hook should only transfer the data and not modify it, nor should it move the data to
+            any other device than the one passed in as argument (unless you know what you are doing).
+            The :class:`~pytorch_lightning.trainer.trainer.Trainer` already takes care of splitting the
+            batch and determines the target devices.
+
+        See Also:
+            - :func:`~pytorch_lightning.utilities.apply_func.move_data_to_device`
+            - :func:`~pytorch_lightning.utilities.apply_func.apply_to_collection`
+        """
+        return move_data_to_device(batch, device)

pytorch_lightning/trainer/distrib_parts.py

@@ -18,6 +18,7 @@
     LightningDistributedDataParallel,
     LightningDataParallel,
 )
+from pytorch_lightning.utilities import move_data_to_device
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.distributed import rank_zero_only
 
@@ -99,58 +100,50 @@ def copy_trainer_model_properties(self, model):
             m.tpu_local_core_rank = self.tpu_local_core_rank
             m.tpu_global_core_rank = self.tpu_global_core_rank
 
-    def transfer_batch_to_tpu(self, batch):
-        return self.__transfer_data_to_device(batch, device='tpu')
-
-    def transfer_batch_to_gpu(self, batch, gpu_id):
-        return self.__transfer_data_to_device(batch, device='gpu', gpu_id=gpu_id)
-
-    def __transfer_data_to_device(self, batch, device, gpu_id=None):
-        if device == 'tpu' and XLA_AVAILABLE:
-            # base case: object can be directly moved using `to`
-            if callable(getattr(batch, 'to', None)):
-                xla_device = xm.xla_device(self.tpu_id) if self.tpu_id is not None else xm.xla_device()
-                return batch.to(xla_device)
-
-        if device == 'gpu':
-            # base case: object can be directly moved using `cuda` or `to`
-            if callable(getattr(batch, 'cuda', None)):
-                # non_blocking will be ignored if tensor is not pinned.
-                # so we can always set it to True
-                return batch.cuda(gpu_id, non_blocking=True)
-
-            if callable(getattr(batch, 'to', None)):
-                # non_blocking will be ignored if tensor is not pinned.
-                # so we can always set it to True
-                return batch.to(torch.device('cuda', gpu_id), non_blocking=True)
-
-        # when list
-        if isinstance(batch, list):
-            for i, x in enumerate(batch):
-                batch[i] = self.__transfer_data_to_device(x, device, gpu_id)
-            return batch
-
-        # when tuple
-        if isinstance(batch, tuple):
-            # when namedtuple
-            if hasattr(batch, '_fields'):
-                elem_type = type(batch)
-                return elem_type(*(self.__transfer_data_to_device(x, device, gpu_id) for x in batch))
-            else:
-                batch = list(batch)
-                for i, x in enumerate(batch):
-                    batch[i] = self.__transfer_data_to_device(x, device, gpu_id)
-                return tuple(batch)
-
-        # when dict
-        if isinstance(batch, dict):
-            for k, v in batch.items():
-                batch[k] = self.__transfer_data_to_device(v, device, gpu_id)
-
-            return batch
-
-        # nothing matches, return the value as is without transform
-        return batch
+    def transfer_batch_to_tpu(self, batch: Any, tpu_id: Optional[int] = None):
+        """
+        Transfers the data to the TPU.
+
+        Args:
+            batch: A tensor or collection of tensors.
+            tpu_id: The id of the TPU core. If omitted, the first available core is chosen.
+
+        Return:
+            the tensor on the TPU device.
+
+        See Also:
+            - :func:`~pytorch_lightning.utilities.apply_func.move_data_to_device`
+        """
+        if not XLA_AVAILABLE:
+            raise MisconfigurationException(
+                'Requested to transfer batch to TPU but XLA is not available.'
+                ' Are you sure this machine has TPUs?'
+            )
+        device = xm.xla_device(tpu_id)
+        return self.__transfer_batch_to_device(batch, device)
+
+    def transfer_batch_to_gpu(self, batch: Any, gpu_id: Optional[int] = None):
+        """
+        Transfers the data to the GPU.
+
+        Args:
+            batch: A tensor or collection of tensors.
+            gpu_id: The id of the GPU device. If omitted, the first available GPU is chosen.
+
+        Return:
+            the tensor on the GPU device.
+
+        See Also:
+            - :func:`~pytorch_lightning.utilities.apply_func.move_data_to_device`
+        """
+        device = torch.device('cuda', gpu_id)
+        return self.__transfer_batch_to_device(batch, device)
+
+    def __transfer_batch_to_device(self, batch: Any, device: torch.device):
+        model = self.get_model()
+        if model is not None:
+            return model.transfer_batch_to_device(batch, device)
+        return move_data_to_device(batch, device)
 
     def single_gpu_train(self, model):
         model.cuda(self.root_gpu)

pytorch_lightning/trainer/evaluation_loop.py

@@ -434,7 +434,7 @@ def evaluation_forward(self, model, batch, batch_idx, dataloader_idx, test_mode:
 
         # TPU data  transfer
         if self.use_tpu:
-            batch = self.transfer_batch_to_tpu(batch)
+            batch = self.transfer_batch_to_tpu(batch, self.tpu_id)
             args[0] = batch
 
         # CPU, TPU or gpu step

pytorch_lightning/trainer/training_loop.py

@@ -753,7 +753,7 @@ def training_forward(self, batch, batch_idx, opt_idx, hiddens):
 
         # TPU support
         elif self.use_tpu:
-            batch = self.transfer_batch_to_tpu(batch)
+            batch = self.transfer_batch_to_tpu(batch, self.tpu_id)
             args[0] = batch
             output = self.model.training_step(*args)
 

pytorch_lightning/utilities/__init__.py

@@ -1,3 +1,4 @@
 """General utilities"""
 
 from pytorch_lightning.utilities.distributed import rank_zero_only, rank_zero_warn
+from pytorch_lightning.utilities.apply_func import move_data_to_device

pytorch_lightning/utilities/apply_func.py

@@ -1,6 +1,8 @@
 from collections import Mapping, Sequence
 from typing import Any, Callable, Union
 
+import torch
+
 
 def apply_to_collection(data: Any, dtype: Union[type, tuple], function: Callable, *args, **kwargs) -> Any:
     """
@@ -34,3 +36,24 @@ def apply_to_collection(data: Any, dtype: Union[type, tuple], function: Callable
 
     # data is neither of dtype, nor a collection
     return data
+
+
+def move_data_to_device(batch: Any, device: torch.device):
+    """
+    Transfers a collection of tensors to the given device.
+
+    Args:
+        batch: A tensor or collection of tensors. See :func:`apply_to_collection`
+            for a list of supported collection types.
+        device: The device to which tensors should be moved
+
+    Return:
+        the same collection but with all contained tensors residing on the new device.
+
+    See Also:
+        - :meth:`torch.Tensor.to`
+        - :class:`torch.device`
+    """
+    def to(tensor):
+        return tensor.to(device, non_blocking=True)
+    return apply_to_collection(batch, dtype=torch.Tensor, function=to)

tests/models/test_hooks.py

@@ -1,3 +1,5 @@
+from unittest.mock import MagicMock
+
 import pytest
 import torch
 
@@ -68,3 +70,37 @@ def training_epoch_end(self, outputs):
     # metrics are kept after each epoch
     for i in range(num_epochs):
         assert metrics[f'epoch_metric_{i}'] == i
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="test requires GPU machine")
+def test_transfer_batch_hook():
+
+    class CustomBatch:
+
+        def __init__(self, data):
+            self.samples = data[0]
+            self.targets = data[1]
+
+    class CurrentTestModel(EvalModelTemplate):
+
+        hook_called = False
+
+        def transfer_batch_to_device(self, data, device):
+            self.hook_called = True
+            if isinstance(data, CustomBatch):
+                data.samples = data.samples.to(device)
+                data.targets = data.targets.to(device)
+            else:
+                data = super().transfer_batch_to_device(data, device)
+            return data
+
+    model = CurrentTestModel()
+    batch = CustomBatch((torch.zeros(5, 28), torch.ones(5, 1, dtype=torch.long)))
+
+    trainer = Trainer()
+    # running .fit() would require us to implement custom data loaders, we mock the model reference instead
+    trainer.get_model = MagicMock(return_value=model)
+    batch_gpu = trainer.transfer_batch_to_gpu(batch, 0)
+    expected = torch.device('cuda', 0)
+    assert model.hook_called
+    assert batch_gpu.samples.device == batch_gpu.targets.device == expected