copy_batch_to_device functionality (#175)

ocf_data_sampler/sample/base.py

@@ -73,3 +73,26 @@ def batch_to_tensor(batch: NumpyBatch) -> TensorBatch:
             elif np.issubdtype(v.dtype, np.number):
                 batch[k] = torch.as_tensor(v)
     return batch
+
+
+def copy_batch_to_device(batch: dict, device: torch.device) -> dict:
+    """
+    Moves tensor leaves in a nested dict to a new device.
+
+    Args:
+        batch: Nested dict with tensors to move.
+        device: Device to move tensors to.
+
+    Returns:
+        A dict with tensors moved to the new device.
+    """
+    batch_copy = {}
+
+    for k, v in batch.items():
+        if isinstance(v, dict):
+            batch_copy[k] = copy_batch_to_device(v, device)  
+        elif isinstance(v, torch.Tensor):
+            batch_copy[k] = v.to(device)
+        else:
+            batch_copy[k] = v
+    return batch_copy

scripts/refactor_site.py

@@ -1,5 +1,5 @@
 """ Helper functions for refactoring legacy site data """
-
+import xarray as xr
 
 def legacy_format(data_ds, metadata_df):
     """This formats old legacy data to the new format.

tests/test_sample/test_base.py

@@ -9,7 +9,8 @@
 from pathlib import Path
 from ocf_data_sampler.sample.base import (
     SampleBase,
-    batch_to_tensor
+    batch_to_tensor,
+    copy_batch_to_device
 )
 
 class TestSample(SampleBase):
@@ -145,3 +146,19 @@ def test_batch_to_tensor_multidimensional():
     assert tensor_batch['matrix'].shape == (2, 2)
     assert tensor_batch['tensor'].shape == (2, 2, 2)
     assert torch.equal(tensor_batch['matrix'], torch.tensor([[1, 2], [3, 4]]))
+
+
+def test_copy_batch_to_device():
+    """ Test moving tensors to a different device """
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    batch = {
+        'tensor_data': torch.tensor([1, 2, 3]),
+        'nested': {
+            'matrix': torch.tensor([[1, 2], [3, 4]])
+        },
+        'non_tensor': 'unchanged'
+    }
+    moved_batch = copy_batch_to_device(batch, device)
+    assert moved_batch['tensor_data'].device == device
+    assert moved_batch['nested']['matrix'].device == device
+    assert moved_batch['non_tensor'] == 'unchanged'  # Non-tensors should remain unchanged