perf: skip redundant convert_to_tensor for MetaTensor strip

Several transforms invoke ``convert_to_tensor`` twice on the same input:
once with ``track_meta=get_track_meta()`` to keep the MetaTensor wrapper,
then again with ``track_meta=False`` to obtain a plain-tensor view for the
actual computation. The second call routes through
``_convert_tensor(data).to(dtype=..., device=..., memory_format=contiguous_format)``
(``monai/utils/type_conversion.py``) and is functionally equivalent to
calling ``MetaTensor.as_tensor()`` for the purpose intended here, since
the first call already enforced ``contiguous_format``.

Replace the second call with an explicit ``as_tensor()`` strip:

    img_t = img.as_tensor() if isinstance(img, MetaTensor) else img

This avoids a redundant ``convert_to_tensor`` dispatch and a no-op
``Tensor.to(memory_format=contiguous_format)`` per ``__call__``. Sites
touched (all in transform ``__call__``, so per-sample per-epoch):

* ``KeepLargestConnectedComponent`` (post/array.py)
* ``RemoveSmallObjects`` (post/array.py)
* ``LabelToContour`` (post/array.py)
* ``ScaleIntensity`` (intensity/array.py)
* ``ScaleIntensityFixedMean`` (intensity/array.py)
* ``ClipIntensityPercentiles`` (intensity/array.py)
* ``NormalizeIntensity`` (intensity/array.py)
* ``SavitzkyGolaySmooth`` (intensity/array.py)
* ``RandHistogramShift`` (intensity/array.py)
* ``KSpaceSpikeNoise`` (intensity/array.py)
* ``IntensityRemap`` (intensity/array.py)

No behavioral change: the resulting tensor is the same underlying data
(``as_tensor()`` returns the wrapped ``torch.Tensor`` without copying).

Signed-off-by: Soumya Snigdha Kundu <soumya_snigdha.kundu@kcl.ac.uk>

Author: aymuos15

monai/transforms/intensity/array.py

@@ -26,6 +26,7 @@
 from monai.config import DtypeLike
 from monai.config.type_definitions import NdarrayOrTensor, NdarrayTensor
 from monai.data.meta_obj import get_track_meta
+from monai.data.meta_tensor import MetaTensor
 from monai.data.ultrasound_confidence_map import UltrasoundConfidenceMap
 from monai.data.utils import get_random_patch, get_valid_patch_size
 from monai.networks.layers import GaussianFilter, HilbertTransform, MedianFilter, SavitzkyGolayFilter
@@ -483,7 +484,7 @@ def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
 
         """
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_t = convert_to_tensor(img, track_meta=False)
+        img_t = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         ret: NdarrayOrTensor
         if self.minv is not None or self.maxv is not None:
             if self.channel_wise:
@@ -542,7 +543,7 @@ def __call__(self, img: NdarrayOrTensor, factor=None) -> NdarrayOrTensor:
         factor = factor if factor is not None else self.factor
 
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_t = convert_to_tensor(img, track_meta=False)
+        img_t = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         ret: NdarrayOrTensor
         if self.channel_wise:
             out = []
@@ -1168,7 +1169,7 @@ def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
         Apply the transform to `img`.
         """
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_t = convert_to_tensor(img, track_meta=False)
+        img_t = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         if self.channel_wise:
             img_t = torch.stack([self._clip(img=d) for d in img_t])  # type: ignore
         else:
@@ -1433,7 +1434,7 @@ def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
         Apply the transform to `img`.
         """
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_t = convert_to_tensor(img, track_meta=False)
+        img_t = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         if self.channel_wise:
             img_t = torch.stack([self._normalize(img=d) for d in img_t])  # type: ignore
         else:
@@ -1530,7 +1531,7 @@ def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
 
         """
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        self.img_t = convert_to_tensor(img, track_meta=False)
+        self.img_t = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
 
         # add one to transform axis because a batch axis will be added at dimension 0
         savgol_filter = SavitzkyGolayFilter(self.window_length, self.order, self.axis + 1, self.mode)
@@ -1907,7 +1908,7 @@ def __call__(self, img: NdarrayOrTensor, randomize: bool = True) -> NdarrayOrTen
 
         if self.reference_control_points is None or self.floating_control_points is None:
             raise RuntimeError("please call the `randomize()` function first.")
-        img_t = convert_to_tensor(img, track_meta=False)
+        img_t = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         img_min, img_max = img_t.min(), img_t.max()
         if img_min == img_max:
             warn(
@@ -1952,7 +1953,7 @@ def __init__(self, alpha: float = 0.1) -> None:
 
     def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_t = convert_to_tensor(img, track_meta=False)
+        img_t = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         n_dims = len(img_t.shape[1:])
 
         # FT
@@ -2604,7 +2605,7 @@ def __call__(self, img: torch.Tensor) -> torch.Tensor:
             img: image to remap.
         """
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_ = convert_to_tensor(img, track_meta=False)
+        img_ = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         # sample noise
         vals_to_sample = torch.unique(img_).tolist()
         noise = torch.from_numpy(self.R.choice(vals_to_sample, len(vals_to_sample) - 1 + self.kernel_size))

monai/transforms/post/array.py

@@ -16,6 +16,7 @@
 
 import warnings
 from collections.abc import Callable, Iterable, Sequence
+from typing import cast
 
 import numpy as np
 import torch
@@ -338,7 +339,7 @@ def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
         else:
             applied_labels = tuple(get_unique_labels(img, is_onehot, discard=0))
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_: torch.Tensor = convert_to_tensor(img, track_meta=False)
+        img_: torch.Tensor = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         if self.independent:
             for i in applied_labels:
                 foreground = img_[i] > 0 if is_onehot else img_[0] == i
@@ -497,7 +498,7 @@ def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
 
         if isinstance(img, torch.Tensor):
             img = convert_to_tensor(img, track_meta=get_track_meta())
-            img_ = convert_to_tensor(img, track_meta=False)
+            img_ = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
             if hasattr(torch, "isin"):  # `isin` is new in torch 1.10.0
                 appl_lbls = torch.as_tensor(self.applied_labels, device=img_.device)
                 out = torch.where(torch.isin(img_, appl_lbls), img_, torch.tensor(0.0).to(img_))
@@ -623,7 +624,7 @@ def __call__(self, img: NdarrayOrTensor) -> NdarrayOrTensor:
 
         """
         img = convert_to_tensor(img, track_meta=get_track_meta())
-        img_: torch.Tensor = convert_to_tensor(img, track_meta=False)
+        img_: torch.Tensor = img.as_tensor() if isinstance(img, MetaTensor) else cast(torch.Tensor, img)
         spatial_dims = len(img_.shape) - 1
         img_ = img_.unsqueeze(0)  # adds a batch dim
         if spatial_dims == 2: