Speedups & fixes

changelog.md

@@ -1,3 +1,9 @@
+# v0.3.4
+
+- Fix a data_dims access issue
+- Marginally improve the speed of handling FoldedTensors in standard torch operations
+- Use default torch types (e.g. `torch.float32` or `torch.torch64`)
+
 # v0.3.3
 
 - Handle empty inputs (e.g. `as_folded_tensor([[[], []], [[]]])`) by returning an empty tensor

foldedtensor/__init__.py

@@ -5,6 +5,8 @@
 import torch
 from torch.autograd import Function
 
+from . import _C
+
 np_to_torch_dtype = {
     torch.bool: bool,
     torch.uint8: np.uint8,
@@ -19,9 +21,29 @@
     torch.complex128: np.complex128,
 }
 
-from . import _C
+pass_through_functions = {
+    torch.Tensor._grad.__get__,
+    torch.Tensor.grad,
+    torch.Tensor._base.__get__,
+    torch.Tensor.__repr__,
+    torch.Tensor.__str__,
+    torch.Tensor.__format__,
+    torch.Tensor.shape.__get__,
+    torch.Tensor.size.__get__,
+    torch.Tensor.dtype.__get__,
+    torch.Tensor.device.__get__,
+}
+if hasattr(torch._C, "TensorBase"):
+    pass_through_functions.add(torch._C.TensorBase.size)
+else:
+    pass_through_functions.add(torch.Tensor.size)
+
+try:
+    DisableTorchFunctionSubclass = torch._C.DisableTorchFunctionSubclass
+except AttributeError:
+    DisableTorchFunctionSubclass = torch._C.DisableTorchFunction
 
-__version__ = "0.3.3"
+__version__ = "0.3.4"
 
 
 # noinspection PyMethodOverriding
@@ -71,7 +93,6 @@ def backward(ctx, grad_output):
             ctx.lengths,
             ctx.old_data_dims,
         )
-        # new_data_flat.index_put_({new_indexer}, old_data_flat.index_select(0, old_indexer));
         shape_suffix = grad_output.shape[len(ctx.new_data_dims) :]
         grad_input = torch.zeros(
             (*shape_prefix, *shape_suffix), dtype=grad_output.dtype, device=device
@@ -80,20 +101,13 @@ def backward(ctx, grad_output):
             -1, *shape_suffix
         ).index_select(0, ctx.output_indexer)
         return grad_input, None
-        # return FoldedTensor(
-        #     data=refolded_data,
-        #     lengths=ctx.lengths,
-        #     data_dims=ctx.old_data_dims,
-        #     full_names=full_names,
-        #     indexer=indexer,
-        # )
 
 
 type_to_dtype_dict = {
-    int: torch.int64,
-    float: torch.float64,
+    int: torch.tensor([0]).dtype,
+    float: torch.tensor([0.0]).dtype,
     bool: torch.bool,
-    None: torch.float64,
+    None: torch.tensor([0.0]).dtype,
 }
 
 
@@ -151,16 +165,18 @@ def as_folded_tensor(
             )
             if (data_dims[-1] + 1) != len(full_names):
                 raise ValueError(
-                    "The last dimension of `data_dims` must be the last variable dimension."
+                    "The last dimension of `data_dims` must be the last "
+                    "variable dimension."
                 )
         elif full_names is not None:
             data_dims = tuple(range(len(full_names)))
     if isinstance(data, torch.Tensor) and lengths is not None:
         data_dims = data_dims or tuple(range(len(lengths)))
         np_indexer, shape = _C.make_refolding_indexer(lengths, data_dims)
-        assert shape == list(
-            data.shape[: len(data_dims)]
-        ), f"Shape inferred from lengths is not compatible with data dims: {shape}, {data.shape}, {len(data_dims)}"
+        assert shape == list(data.shape[: len(data_dims)]), (
+            f"Shape inferred from lengths is not compatible with data dims: {shape}, "
+            f"{data.shape}, {len(data_dims)}"
+        )
         result = FoldedTensor(
             data=data,
             lengths=lengths,
@@ -208,6 +224,23 @@ def as_folded_tensor(
     return result
 
 
+def _postprocess_func_result(result, input):
+    if (
+        input is not None
+        and input.shape[: len(input.data_dims)] != result.shape[: len(input.data_dims)]
+    ):
+        return result
+
+    return FoldedTensor(
+        data=result,
+        lengths=input.lengths,
+        data_dims=input.data_dims,
+        full_names=input.full_names,
+        indexer=input.indexer,
+        mask=input._mask,
+    )
+
+
 # noinspection PyUnresolvedReferences,PyInitNewSignature
 class FoldedTensor(torch.Tensor):
     """
@@ -296,46 +329,52 @@ def to(self, *args, **kwargs):
 
     @classmethod
     def __torch_function__(cls, func, types, args=(), kwargs=None):
-        result = super().__torch_function__(func, types, args, kwargs)
+        if kwargs is None:
+            kwargs = {}
+        if func in pass_through_functions:
+            with DisableTorchFunctionSubclass():
+                return func(*args, **kwargs)
+        with DisableTorchFunctionSubclass():
+            result = func(*args, **kwargs)
 
         if func is torch.Tensor.share_memory_:
             self = args[0]
             self.indexer.share_memory_()
             if self._mask is not None:
                 self._mask.share_memory_()
-
-        if not isinstance(result, torch.Tensor):
-            return result
+            return self
 
         ft = None
-        for arg in (*args, *(kwargs or {}).values()):
+        for arg in (*args, *kwargs.values()):
             if isinstance(arg, FoldedTensor):
                 assert (
                     ft is None or ft.data_dims == arg.data_dims
-                ), "Cannot perform operation on FoldedTensors with different structure"
+                ), "Cannot perform operation on FoldedTensors with different structures"
                 ft = arg
-            if isinstance(arg, (list, tuple)):
+            elif isinstance(arg, (list, tuple)):
                 for item in arg:
                     if isinstance(item, FoldedTensor):
-                        assert (
-                            ft is None or ft.data_dims == item.data_dims
-                        ), "Cannot perform operation on FoldedTensors with different structure"
+                        assert ft is None or ft.data_dims == item.data_dims, (
+                            "Cannot perform operation on FoldedTensors with "
+                            "different structures"
+                        )
                         ft = item
 
+        if isinstance(result, torch.Tensor):
+            return _postprocess_func_result(result, ft)
+
         if (
-            ft is not None
-            and ft.shape[: len(ft.data_dims)] != result.shape[: len(ft.data_dims)]
+            isinstance(result, (tuple, list))
+            and len(result)
+            and isinstance(result[0], torch.Tensor)
         ):
-            return result.as_subclass(torch.Tensor)
+            return type(result)(
+                _postprocess_func_result(item, ft)
+                if isinstance(item, FoldedTensor)
+                else item
+                for item in result
+            )
 
-        result = FoldedTensor(
-            data=result,
-            lengths=ft.lengths,
-            data_dims=ft.data_dims,
-            full_names=ft.full_names,
-            indexer=ft.indexer,
-            mask=ft._mask,
-        )
         return result
 
     def clone(self):

pyproject.toml

@@ -9,7 +9,7 @@ readme = "README.md"
 urls.homepage = "https://github.com/aphp/foldedtensor/"
 urls.repository = "https://github.com/aphp/foldedtensor/"
 dynamic = ["version"]
-requires-python = ">3.7.6,<4.0"
+requires-python = ">3.7.1,<4.0"
 
 dependencies = [
     "torch>1.0.0",
@@ -82,7 +82,6 @@ fix = true
 exclude = [
     ".git",
     "__pycache__",
-    "__init__.py",
     ".mypy_cache",
     ".pytest_cache",
     ".venv",
@@ -98,5 +97,6 @@ select = [
 ]
 fixable = ["E", "F", "W", "I"]
 
-[isort]
-known-first-party = ["foldedtensor"]
+[tool.ruff.isort]
+    known-first-party = ["foldedtensor"]
+    known-third-party = ["build"]

tests/test_folded_tensor.py

@@ -404,3 +404,16 @@ def test_imbalanced_sequence_2():
         )
 
     assert "'int' object is not iterable" in str(e.value)
+
+
+def test_max():
+    ft = as_folded_tensor(
+        [
+            [0, 1, 2],
+            [3, 4],
+        ],
+        dtype=torch.float,
+    )
+    values, indices = ft.max(-1)
+    assert (values == torch.tensor([2, 4])).all()
+    assert (indices == torch.tensor([2, 1])).all()