Allow for specifying a tensor device in AnnLoader

anndata/experimental/pytorch/_annloader.py

@@ -4,6 +4,7 @@
 from functools import partial
 from math import ceil
 from typing import TYPE_CHECKING
+from warnings import warn
 
 import numpy as np
 from scipy.sparse import issparse
@@ -54,18 +55,17 @@ def __len__(self):
         return length
 
 
-# maybe replace use_cuda with explicit device option
-def default_converter(arr, use_cuda, pin_memory):
+def default_converter(arr, device, pin_memory):
     if isinstance(arr, torch.Tensor):
-        if use_cuda:
-            arr = arr.cuda()
+        if device is not None:
+            arr = arr.to(device)
         elif pin_memory:
             arr = arr.pin_memory()
     elif arr.dtype.name != "category" and np.issubdtype(arr.dtype, np.number):
         if issparse(arr):
             arr = arr.toarray()
-        if use_cuda:
-            arr = torch.tensor(arr, device="cuda")
+        if device is not None:
+            arr = torch.tensor(arr, device=device)
         else:
             arr = torch.tensor(arr)
             arr = arr.pin_memory() if pin_memory else arr
@@ -114,12 +114,15 @@ class AnnLoader(DataLoader):
         Set to `True` to have the data reshuffled at every epoch.
     use_default_converter
         Use the default converter to convert arrays to pytorch tensors, transfer to
-        the default cuda device (if `use_cuda=True`), do memory pinning (if `pin_memory=True`).
+        the specified device (if `device!=None`), do memory pinning (if `pin_memory=True`).
         If you pass an AnnCollection object with prespecified converters, the default converter
         won't overwrite these converters but will be applied on top of them.
     use_cuda
         Transfer pytorch tensors to the default cuda device after conversion.
-        Only works if `use_default_converter=True`
+        Only works if `use_default_converter=True`. DEPRECATED in favour of `device`.
+    device
+        The device to which to transfer pytorch tensors after conversion (example: "cuda").
+        Only works if `use_default_converter=True`.
     **kwargs
         Arguments for PyTorch DataLoader. If `adatas` is not an `AnnCollection` object, then also
         arguments for `AnnCollection` initialization.
@@ -132,8 +135,16 @@ def __init__(
         shuffle: bool = False,
         use_default_converter: bool = True,
         use_cuda: bool = False,
+        device: str | None = None,
         **kwargs,
     ):
+        if use_cuda:
+            warn(
+                "Argument use_cuda has been deprecated in favour of `device`. ",
+                FutureWarning,
+            )
+            device = "cuda"
+
         if isinstance(adatas, AnnData):
             adatas = [adatas]
 
@@ -171,7 +182,7 @@ def __init__(
         if use_default_converter:
             pin_memory = kwargs.pop("pin_memory", False)
             _converter = partial(
-                default_converter, use_cuda=use_cuda, pin_memory=pin_memory
+                default_converter, device=device, pin_memory=pin_memory
             )
             dataset.convert = _convert_on_top(
                 dataset.convert, _converter, dict(dataset.attrs_keys, X=[])

docs/release-notes/0.10.4.md

@@ -1,5 +1,9 @@
 ### 0.10.4 {small}`the future`
 
+```{rubric} New features
+```
+* `AnnLoader` now accepts a `device` argument to specify the device to load the data to {pr}`1240` {user}`austinv11`
+
 ```{rubric} Bugfix
 ```
 * Only try to use `Categorical.map(na_action=…)` in actually supported Pandas ≥2.1 {pr}`1226` {user}`flying-sheep`
@@ -10,3 +14,7 @@
 
 ```{rubric} Performance
 ```
+
+```{rubric} Deprecations
+```
+* `AnnLoader(use_cuda=…)` is deprecated in favour of `AnnLoader(device=…)` {pr}`1240` {user}`austinv11