Merge branch 'dev' into fix-loadimage-reader-exception

Author: ericspod

docs/requirements.txt

@@ -40,6 +40,6 @@ onnx>=1.13.0
 onnxruntime; python_version <= '3.10'
 zarr
 huggingface_hub
-pyamg>=5.0.0
+pyamg>=5.0.0, <5.3.0
 packaging
 polygraphy

monai/inferers/inferer.py

@@ -916,6 +916,7 @@ def sample(
         verbose: bool = True,
         seg: torch.Tensor | None = None,
         cfg: float | None = None,
+        cfg_fill_value: float = -1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, list[torch.Tensor]]:
         """
         Args:
@@ -929,6 +930,7 @@ def sample(
             verbose: if true, prints the progression bar of the sampling process.
             seg: if diffusion model is instance of SPADEDiffusionModel, segmentation must be provided.
             cfg: classifier-free-guidance scale, which indicates the level of strengthening on the conditioning.
+            cfg_fill_value: the fill value to use for the unconditioned input when using classifier-free guidance.
         """
         if mode not in ["crossattn", "concat"]:
             raise NotImplementedError(f"{mode} condition is not supported")
@@ -961,7 +963,7 @@ def sample(
                 model_input = torch.cat([image] * 2, dim=0)
                 if conditioning is not None:
                     uncondition = torch.ones_like(conditioning)
-                    uncondition.fill_(-1)
+                    uncondition.fill_(cfg_fill_value)
                     conditioning_input = torch.cat([uncondition, conditioning], dim=0)
                 else:
                     conditioning_input = None
@@ -1261,6 +1263,7 @@ def sample(  # type: ignore[override]
         verbose: bool = True,
         seg: torch.Tensor | None = None,
         cfg: float | None = None,
+        cfg_fill_value: float = -1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, list[torch.Tensor]]:
         """
         Args:
@@ -1276,6 +1279,7 @@ def sample(  # type: ignore[override]
             seg: if diffusion model is instance of SPADEDiffusionModel, or autoencoder_model
              is instance of SPADEAutoencoderKL, segmentation must be provided.
             cfg: classifier-free-guidance scale, which indicates the level of strengthening on the conditioning.
+            cfg_fill_value: the fill value to use for the unconditioned input when using classifier-free guidance.
         """
 
         if (
@@ -1300,6 +1304,7 @@ def sample(  # type: ignore[override]
             verbose=verbose,
             seg=seg,
             cfg=cfg,
+            cfg_fill_value=cfg_fill_value,
         )
 
         if save_intermediates:
@@ -1479,6 +1484,7 @@ def sample(  # type: ignore[override]
         verbose: bool = True,
         seg: torch.Tensor | None = None,
         cfg: float | None = None,
+        cfg_fill_value: float = -1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, list[torch.Tensor]]:
         """
         Args:
@@ -1493,7 +1499,8 @@ def sample(  # type: ignore[override]
             mode: Conditioning mode for the network.
             verbose: if true, prints the progression bar of the sampling process.
             seg: if diffusion model is instance of SPADEDiffusionModel, segmentation must be provided.
-                        cfg: classifier-free-guidance scale, which indicates the level of strengthening on the conditioning.
+            cfg: classifier-free-guidance scale, which indicates the level of strengthening on the conditioning.
+            cfg_fill_value: the fill value to use for the unconditioned input when using classifier-free guidance.
         """
         if mode not in ["crossattn", "concat"]:
             raise NotImplementedError(f"{mode} condition is not supported")
@@ -1521,7 +1528,7 @@ def sample(  # type: ignore[override]
                 model_input = torch.cat([image] * 2, dim=0)
                 if conditioning is not None:
                     uncondition = torch.ones_like(conditioning)
-                    uncondition.fill_(-1)
+                    uncondition.fill_(cfg_fill_value)
                     conditioning_input = torch.cat([uncondition, conditioning], dim=0)
                 else:
                     conditioning_input = None
@@ -1839,6 +1846,7 @@ def sample(  # type: ignore[override]
         verbose: bool = True,
         seg: torch.Tensor | None = None,
         cfg: float | None = None,
+        cfg_fill_value: float = -1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, list[torch.Tensor]]:
         """
         Args:
@@ -1856,6 +1864,7 @@ def sample(  # type: ignore[override]
             seg: if diffusion model is instance of SPADEDiffusionModel, or autoencoder_model
              is instance of SPADEAutoencoderKL, segmentation must be provided.
             cfg: classifier-free-guidance scale, which indicates the level of strengthening on the conditioning.
+            cfg_fill_value: the fill value to use for the unconditioned input when using classifier-free guidance.
         """
 
         if (
@@ -1884,6 +1893,7 @@ def sample(  # type: ignore[override]
             verbose=verbose,
             seg=seg,
             cfg=cfg,
+            cfg_fill_value=cfg_fill_value,
         )
 
         if save_intermediates:

monai/transforms/spatial/array.py

@@ -64,6 +64,7 @@
     GridSamplePadMode,
     InterpolateMode,
     NumpyPadMode,
+    SpaceKeys,
     convert_to_cupy,
     convert_to_dst_type,
     convert_to_numpy,
@@ -75,6 +76,7 @@
     issequenceiterable,
     optional_import,
 )
+from monai.utils.deprecate_utils import deprecated_arg_default
 from monai.utils.enums import GridPatchSort, PatchKeys, TraceKeys, TransformBackends
 from monai.utils.misc import ImageMetaKey as Key
 from monai.utils.module import look_up_option
@@ -556,11 +558,20 @@ class Orientation(InvertibleTransform, LazyTransform):
 
     backend = [TransformBackends.NUMPY, TransformBackends.TORCH]
 
+    @deprecated_arg_default(
+        name="labels",
+        old_default=(("L", "R"), ("P", "A"), ("I", "S")),
+        new_default=None,
+        msg_suffix=(
+            "Default value changed to None meaning that the transform now uses the 'space' of a "
+            "meta-tensor, if applicable, to determine appropriate axis labels."
+        ),
+    )
     def __init__(
         self,
         axcodes: str | None = None,
         as_closest_canonical: bool = False,
-        labels: Sequence[tuple[str, str]] | None = (("L", "R"), ("P", "A"), ("I", "S")),
+        labels: Sequence[tuple[str, str]] | None = None,
         lazy: bool = False,
     ) -> None:
         """
@@ -573,7 +584,14 @@ def __init__(
             as_closest_canonical: if True, load the image as closest to canonical axis format.
             labels: optional, None or sequence of (2,) sequences
                 (2,) sequences are labels for (beginning, end) of output axis.
-                Defaults to ``(('L', 'R'), ('P', 'A'), ('I', 'S'))``.
+                If ``None``, an appropriate value is chosen depending on the
+                value of the ``"space"`` metadata item of a metatensor: if
+                ``"space"`` is ``"LPS"``, the value used is ``(('R', 'L'),
+                ('A', 'P'), ('I', 'S'))``, if ``"space"`` is ``"RPS"`` or the
+                input is not a meta-tensor or has no ``"space"`` item, the
+                value ``(('L', 'R'), ('P', 'A'), ('I', 'S'))`` is used. If not
+                ``None``, the provided value is always used and the ``"space"``
+                metadata item (if any) of the input is ignored.
             lazy: a flag to indicate whether this transform should execute lazily or not.
                 Defaults to False
 
@@ -619,9 +637,19 @@ def __call__(self, data_array: torch.Tensor, lazy: bool | None = None) -> torch.
             raise ValueError(f"data_array must have at least one spatial dimension, got {spatial_shape}.")
         affine_: np.ndarray
         affine_np: np.ndarray
+        labels = self.labels
         if isinstance(data_array, MetaTensor):
             affine_np, *_ = convert_data_type(data_array.peek_pending_affine(), np.ndarray)
             affine_ = to_affine_nd(sr, affine_np)
+
+            # Set up "labels" such that LPS tensors are handled correctly by default
+            if (
+                self.labels is None
+                and "space" in data_array.meta
+                and SpaceKeys(data_array.meta["space"]) == SpaceKeys.LPS
+            ):
+                labels = (("R", "L"), ("A", "P"), ("I", "S"))  # value for LPS
+
         else:
             warnings.warn("`data_array` is not of type `MetaTensor, assuming affine to be identity.")
             # default to identity
@@ -640,7 +668,7 @@ def __call__(self, data_array: torch.Tensor, lazy: bool | None = None) -> torch.
                     f"{self.__class__.__name__}: spatial shape = {spatial_shape}, channels = {data_array.shape[0]},"
                     "please make sure the input is in the channel-first format."
                 )
-            dst = nib.orientations.axcodes2ornt(self.axcodes[:sr], labels=self.labels)
+            dst = nib.orientations.axcodes2ornt(self.axcodes[:sr], labels=labels)
             if len(dst) < sr:
                 raise ValueError(
                     f"axcodes must match data_array spatially, got axcodes={len(self.axcodes)}D data_array={sr}D"
@@ -653,8 +681,19 @@ def inverse(self, data: torch.Tensor) -> torch.Tensor:
         transform = self.pop_transform(data)
         # Create inverse transform
         orig_affine = transform[TraceKeys.EXTRA_INFO]["original_affine"]
-        orig_axcodes = nib.orientations.aff2axcodes(orig_affine)
-        inverse_transform = Orientation(axcodes=orig_axcodes, as_closest_canonical=False, labels=self.labels)
+        labels = self.labels
+
+        # Set up "labels" such that LPS tensors are handled correctly by default
+        if (
+            isinstance(data, MetaTensor)
+            and self.labels is None
+            and "space" in data.meta
+            and SpaceKeys(data.meta["space"]) == SpaceKeys.LPS
+        ):
+            labels = (("R", "L"), ("A", "P"), ("I", "S"))  # value for LPS
+
+        orig_axcodes = nib.orientations.aff2axcodes(orig_affine, labels=labels)
+        inverse_transform = Orientation(axcodes=orig_axcodes, as_closest_canonical=False, labels=labels)
         # Apply inverse
         with inverse_transform.trace_transform(False):
             data = inverse_transform(data)

monai/transforms/spatial/dictionary.py

@@ -71,6 +71,7 @@
     ensure_tuple_rep,
     fall_back_tuple,
 )
+from monai.utils.deprecate_utils import deprecated_arg_default
 from monai.utils.enums import TraceKeys
 from monai.utils.module import optional_import
 
@@ -545,12 +546,21 @@ class Orientationd(MapTransform, InvertibleTransform, LazyTransform):
 
     backend = Orientation.backend
 
+    @deprecated_arg_default(
+        name="labels",
+        old_default=(("L", "R"), ("P", "A"), ("I", "S")),
+        new_default=None,
+        msg_suffix=(
+            "Default value changed to None meaning that the transform now uses the 'space' of a "
+            "meta-tensor, if applicable, to determine appropriate axis labels."
+        ),
+    )
     def __init__(
         self,
         keys: KeysCollection,
         axcodes: str | None = None,
         as_closest_canonical: bool = False,
-        labels: Sequence[tuple[str, str]] | None = (("L", "R"), ("P", "A"), ("I", "S")),
+        labels: Sequence[tuple[str, str]] | None = None,
         allow_missing_keys: bool = False,
         lazy: bool = False,
     ) -> None:
@@ -564,7 +574,14 @@ def __init__(
             as_closest_canonical: if True, load the image as closest to canonical axis format.
             labels: optional, None or sequence of (2,) sequences
                 (2,) sequences are labels for (beginning, end) of output axis.
-                Defaults to ``(('L', 'R'), ('P', 'A'), ('I', 'S'))``.
+                If ``None``, an appropriate value is chosen depending on the
+                value of the ``"space"`` metadata item of a metatensor: if
+                ``"space"`` is ``"LPS"``, the value used is ``(('R', 'L'),
+                ('A', 'P'), ('I', 'S'))``, if ``"space"`` is ``"RPS"`` or the
+                input is not a meta-tensor or has no ``"space"`` item, the
+                value ``(('L', 'R'), ('P', 'A'), ('I', 'S'))`` is used. If not
+                ``None``, the provided value is always used and the ``"space"``
+                metadata item (if any) of the input is ignored.
             allow_missing_keys: don't raise exception if key is missing.
             lazy: a flag to indicate whether this transform should execute lazily or not.
                 Defaults to False

requirements-dev.txt

@@ -59,7 +59,7 @@ zarr
 lpips==0.1.4
 nvidia-ml-py
 huggingface_hub
-pyamg>=5.0.0
+pyamg>=5.0.0, <5.3.0
 git+https://github.com/facebookresearch/segment-anything.git@6fdee8f2727f4506cfbbe553e23b895e27956588
 onnx_graphsurgeon
 polygraphy

setup.cfg

@@ -87,7 +87,7 @@ all =
     lpips==0.1.4
     nvidia-ml-py
     huggingface_hub
-    pyamg>=5.0.0
+    pyamg>=5.0.0, <5.3.0
 nibabel =
     nibabel
 ninja =
@@ -175,7 +175,7 @@ polygraphy =
 huggingface_hub =
     huggingface_hub
 pyamg =
-    pyamg>=5.0.0
+    pyamg>=5.0.0, <5.3.0
 # segment-anything =
 #     segment-anything @ git+https://github.com/facebookresearch/segment-anything@6fdee8f2727f4506cfbbe553e23b895e27956588#egg=segment-anything
 

tests/inferers/test_diffusion_inferer.py

@@ -106,6 +106,7 @@ def test_sample_cfg(self, model_params, input_shape):
             save_intermediates=True,
             intermediate_steps=1,
             cfg=5,
+            cfg_fill_value=-1,
         )
         self.assertEqual(sample.shape, noise.shape)
 

tests/inferers/test_latent_diffusion_inferer.py

@@ -456,6 +456,7 @@ def test_sample_shape_with_cfg(
                     scheduler=scheduler,
                     seg=input_seg,
                     cfg=5,
+                    cfg_fill_value=-1,
                 )
             else:
                 sample = inferer.sample(