kaiko-eva 0.4.0__py3-none-any.whl → 0.4.1__py3-none-any.whl

eva/vision/data/transforms/croppad/rand_crop_by_label_classes.py

@@ -56,19 +56,20 @@ class RandCropByLabelClasses(base.RandomMonaiTransform):
     def set_random_state(self, seed: int) -> None:
         self._rand_crop.set_random_state(seed)
 
-    def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
+    @override
+    def make_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
         mask = next(inpt for inpt in flat_inputs if isinstance(inpt, tv_tensors.Mask))
         self._rand_crop.randomize(label=mask)
         return {}
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_foreground_crops = self._rand_crop(img=inpt, randomize=False)
         return [tv_tensors.wrap(crop, like=inpt) for crop in inpt_foreground_crops]

eva/vision/data/transforms/croppad/rand_crop_by_pos_neg_label.py

@@ -95,19 +95,20 @@ class RandCropByPosNegLabel(base.RandomMonaiTransform):
     def set_random_state(self, seed: int) -> None:
         self._rand_crop.set_random_state(seed)
 
-    def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
+    @override
+    def make_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
         mask = next(inpt for inpt in flat_inputs if isinstance(inpt, tv_tensors.Mask))
         self._rand_crop.randomize(label=mask)
         return {}
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_foreground_crops = self._rand_crop(img=inpt, randomize=False)
         return [tv_tensors.wrap(crop, like=inpt) for crop in inpt_foreground_crops]

eva/vision/data/transforms/croppad/rand_spatial_crop.py

@@ -5,14 +5,14 @@ from typing import Any, Dict, List, Sequence, Tuple
 
 from monai.transforms.croppad import array as monai_croppad_transforms
 from torchvision import tv_tensors
-from torchvision.transforms import v2
 from torchvision.transforms.v2 import _utils as tv_utils
 from typing_extensions import override
 
 from eva.vision.data import tv_tensors as eva_tv_tensors
+from eva.vision.data.transforms import base
 
 
-class RandSpatialCrop(v2.Transform):
+class RandSpatialCrop(base.TorchvisionTransformV2):
     """Crop image with random size or specific size ROI.
 
     It can crop at a random position as center or at the image center.
@@ -62,19 +62,20 @@ class RandSpatialCrop(v2.Transform):
         """Set the random state for the transform."""
         self._rand_spatial_crop.set_random_state(seed)
 
-    def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
+    @override
+    def make_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
         t, h, w = tv_utils.query_chw(flat_inputs)
         self._rand_spatial_crop.randomize((t, h, w))
         return {}
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         slices = self._get_crop_slices()
         inpt_rand_crop = self._cropper(inpt, slices=slices)

eva/vision/data/transforms/croppad/spatial_pad.py

@@ -6,13 +6,13 @@ from typing import Any, Dict, Sequence
 from monai.transforms.croppad import array as monai_croppad_transforms
 from monai.utils.enums import Method, PytorchPadMode
 from torchvision import tv_tensors
-from torchvision.transforms import v2
 from typing_extensions import override
 
 from eva.vision.data import tv_tensors as eva_tv_tensors
+from eva.vision.data.transforms import base
 
 
-class SpatialPad(v2.Transform):
+class SpatialPad(base.TorchvisionTransformV2):
     """Performs padding to the data.
 
     Padding is applied symmetric for all sides or all on one side for each dimension.
@@ -56,12 +56,12 @@ class SpatialPad(v2.Transform):
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_padded = self._spatial_pad(inpt)
         return tv_tensors.wrap(inpt_padded, like=inpt)

eva/vision/data/transforms/intensity/rand_scale_intensity.py

@@ -53,11 +53,11 @@ class RandScaleIntensity(base.RandomMonaiTransform):
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
     def _(self, inpt: tv_tensors.Image, params: Dict[str, Any]) -> Any:
         inpt_scaled = self._rand_scale_intensity(inpt)
         return tv_tensors.wrap(inpt_scaled, like=inpt)

eva/vision/data/transforms/intensity/rand_shift_intensity.py

@@ -49,11 +49,11 @@ class RandShiftIntensity(base.RandomMonaiTransform):
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
     def _(self, inpt: tv_tensors.Image, params: Dict[str, Any]) -> Any:
         inpt_scaled = self._rand_shift_intensity(inpt)
         return tv_tensors.wrap(inpt_scaled, like=inpt)

eva/vision/data/transforms/intensity/scale_intensity_ranged.py

@@ -5,13 +5,13 @@ from typing import Any, Dict, Tuple
 
 from monai.transforms.intensity import array as monai_intensity_transforms
 from torchvision import tv_tensors
-from torchvision.transforms import v2
 from typing_extensions import override
 
 from eva.vision.data import tv_tensors as eva_tv_tensors
+from eva.vision.data.transforms import base
 
 
-class ScaleIntensityRange(v2.Transform):
+class ScaleIntensityRange(base.TorchvisionTransformV2):
     """Intensity scaling transform.
 
     Scaling from [a_min, a_max] to [b_min, b_max] with clip option.
@@ -46,11 +46,11 @@ class ScaleIntensityRange(v2.Transform):
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
     def _(self, inpt: tv_tensors.Image, params: Dict[str, Any]) -> Any:
         inpt_scaled = self._scale_intensity_range(inpt)
         return tv_tensors.wrap(inpt_scaled, like=inpt)

eva/vision/data/transforms/spatial/flip.py

@@ -6,13 +6,13 @@ from typing import Any, Dict, List, Sequence
 import torch
 from monai.transforms.spatial import array as monai_spatial_transforms
 from torchvision import tv_tensors
-from torchvision.transforms import v2
 from typing_extensions import override
 
 from eva.vision.data import tv_tensors as eva_tv_tensors
+from eva.vision.data.transforms import base
 
 
-class RandFlip(v2.Transform):
+class RandFlip(base.TorchvisionTransformV2):
     """Randomly flips the image along axes."""
 
     def __init__(
@@ -45,23 +45,24 @@ class RandFlip(v2.Transform):
         else:
             self._flips = [monai_spatial_transforms.RandFlip(prob=prob, spatial_axis=spatial_axes)]
 
-    def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
+    @override
+    def make_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
         for flip in self._flips:
             flip.randomize(None)
         return {}
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_flipped = self._apply_flips(inpt)
         return tv_tensors.wrap(inpt_flipped, like=inpt)
 
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_flipped = torch.tensor(self._apply_flips(inpt), dtype=torch.long)
         return tv_tensors.wrap(inpt_flipped, like=inpt)

eva/vision/data/transforms/spatial/resize.py

@@ -7,10 +7,11 @@ from torchvision import tv_tensors
 from torchvision.transforms import v2
 from typing_extensions import override
 
+from eva.vision.data.transforms import base
 from eva.vision.data.transforms.spatial import functional
 
 
-class Resize(v2.Transform):
+class Resize(base.TorchvisionTransformV2):
     """Resize transform for images with spatial or byte-based constraints.
 
     This transform provides two mutually exclusive modes of resizing:
@@ -52,11 +53,11 @@ class Resize(v2.Transform):
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Image)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_resized = self.resize_fn(inpt) if self.resize_fn is not None else inpt
         return tv_tensors.wrap(inpt_resized, like=inpt)

eva/vision/data/transforms/spatial/rotate.py

@@ -5,13 +5,13 @@ from typing import Any, Dict, List
 
 from monai.transforms.spatial import array as monai_spatial_transforms
 from torchvision import tv_tensors
-from torchvision.transforms import v2
 from typing_extensions import override
 
 from eva.vision.data import tv_tensors as eva_tv_tensors
+from eva.vision.data.transforms import base
 
 
-class RandRotate90(v2.Transform):
+class RandRotate90(base.TorchvisionTransformV2):
     """Rotate input tensors by 90 degrees."""
 
     def __init__(
@@ -36,18 +36,19 @@ class RandRotate90(v2.Transform):
             prob=prob, max_k=max_k, spatial_axes=spatial_axes
         )
 
-    def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
+    @override
+    def make_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
         self._rotate.randomize()
         return {}
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_rotated = self._rotate(img=inpt, randomize=False)
         return tv_tensors.wrap(inpt_rotated, like=inpt)

eva/vision/data/transforms/spatial/spacing.py

@@ -8,13 +8,13 @@ import torch
 from monai.data import meta_tensor
 from monai.transforms.spatial import array as monai_spatial_transforms
 from torchvision import tv_tensors
-from torchvision.transforms import v2
 from typing_extensions import override
 
 from eva.vision.data import tv_tensors as eva_tv_tensors
+from eva.vision.data.transforms import base
 
 
-class Spacing(v2.Transform):
+class Spacing(base.TorchvisionTransformV2):
     """Resample input image into the specified `pixdim`.
 
     - Expects tensors of shape `[C, T, H, W]`.
@@ -43,7 +43,8 @@ class Spacing(v2.Transform):
         self._spacing = monai_spatial_transforms.Spacing(pixdim=pixdim, recompute_affine=True)
         self._affine = None
 
-    def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
+    @override
+    def make_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
         self._affine = next(
             inpt.affine for inpt in flat_inputs if isinstance(inpt, eva_tv_tensors.Volume)
         )
@@ -51,17 +52,17 @@ class Spacing(v2.Transform):
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(eva_tv_tensors.Volume)
+    @transform.register(eva_tv_tensors.Volume)
     def _(self, inpt: eva_tv_tensors.Volume, params: Dict[str, Any]) -> Any:
         inpt_spacing = self._spacing(inpt.to_meta_tensor(), mode="bilinear")
         if not isinstance(inpt_spacing, meta_tensor.MetaTensor):
             raise ValueError(f"Expected MetaTensor, got {type(inpt_spacing)}")
         return eva_tv_tensors.Volume.from_meta_tensor(inpt_spacing)
 
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_spacing = self._spacing(
             meta_tensor.MetaTensor(inpt, affine=self._affine), mode="nearest"

eva/vision/data/transforms/utility/ensure_channel_first.py

@@ -5,13 +5,13 @@ from typing import Any, Dict
 
 from monai.transforms.utility import array as monai_utility_transforms
 from torchvision import tv_tensors
-from torchvision.transforms import v2
 from typing_extensions import override
 
 from eva.vision.data import tv_tensors as eva_tv_tensors
+from eva.vision.data.transforms import base
 
 
-class EnsureChannelFirst(v2.Transform):
+class EnsureChannelFirst(base.TorchvisionTransformV2):
     """Adjust or add the channel dimension of input data to ensure `channel_first` shape."""
 
     def __init__(
@@ -40,12 +40,12 @@ class EnsureChannelFirst(v2.Transform):
 
     @functools.singledispatchmethod
     @override
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+    def transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return inpt
 
-    @_transform.register(tv_tensors.Image)
-    @_transform.register(eva_tv_tensors.Volume)
-    @_transform.register(tv_tensors.Mask)
+    @transform.register(tv_tensors.Image)
+    @transform.register(eva_tv_tensors.Volume)
+    @transform.register(tv_tensors.Mask)
     def _(self, inpt: Any, params: Dict[str, Any]) -> Any:
         inpt_channel_first = self._ensure_channel_first(inpt)
         return tv_tensors.wrap(inpt_channel_first, like=inpt)

eva/vision/models/networks/backbones/universal/vit.py

@@ -54,6 +54,30 @@ def vit_small_patch16_224_dino(
     )
 
 
+@backbone_registry.register("universal/vit_tiny_patch16_224_random")
+def vit_tiny_patch16_224_random(
+    dynamic_img_size: bool = True, out_indices: int | Tuple[int, ...] | None = None
+) -> nn.Module:
+    """Initializes a ViT-Tiny16 baseline model with random weights.
+
+    Args:
+        dynamic_img_size: Support different input image sizes by allowing to change
+            the grid size (interpolate abs and/or ROPE pos) in the forward pass.
+        out_indices: Whether and which multi-level patch embeddings to return.
+
+    Returns:
+        The torch ViTS-16 based foundation model.
+    """
+    return timm.create_model(
+        model_name="vit_tiny_patch16_224",
+        pretrained=False,
+        num_classes=0,
+        features_only=out_indices is not None,
+        out_indices=out_indices,
+        dynamic_img_size=dynamic_img_size,
+    )
+
+
 @backbone_registry.register("universal/vit_small_patch16_224_dino_1chan")
 def vit_small_patch16_224_dino_1chan(
     dynamic_img_size: bool = True, out_indices: int | Tuple[int, ...] | None = None

{kaiko_eva-0.4.0.dist-info → kaiko_eva-0.4.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: kaiko-eva
-Version: 0.4.0
+Version: 0.4.1
 Summary: Evaluation Framework for oncology foundation models.
 Keywords: machine-learning,evaluation-framework,oncology,foundation-models
 Author-Email: Ioannis Gatopoulos <ioannis@kaiko.ai>, =?utf-8?q?Nicolas_K=C3=A4nzig?= <nicolas@kaiko.ai>, Roman Moser <roman@kaiko.ai>
@@ -491,7 +491,13 @@ and [tutorials](https://kaiko-ai.github.io/eva/main/user-guide/advanced/replicat
 
 The following table shows the FMs we have evaluated with _`eva`_. For more detailed information about the evaluation process, please refer to our [documentation](https://kaiko-ai.github.io/eva/main/leaderboards/).
 
-![Pathology Leaderboard](./docs/images/leaderboard.svg)
+### Pathology
+
+<img src="./docs/images/leaderboards/pathology.svg" alt="Pathology Leaderboard">
+
+### Radiology
+
+<img src="./docs/images/leaderboards/radiology.svg" alt="Radiology Leaderboard" width="600">
 
 
 ## Contributing