Merge branch 'dev' into dev

Author: ericspod

monai/apps/pathology/transforms/stain/array.py

@@ -85,7 +85,12 @@ def _deconvolution_extract_stain(self, image: np.ndarray) -> np.ndarray:
         v_max = eigvecs[:, 1:3].dot(np.array([(np.cos(max_phi), np.sin(max_phi))], dtype=np.float32).T)
 
         # a heuristic to make the vector corresponding to hematoxylin first and the one corresponding to eosin second
-        if v_min[0] > v_max[0]:
+        # Hematoxylin: high blue, lower red (low R/B ratio)
+        # Eosin: high red, lower blue (high R/B ratio)
+        eps = np.finfo(np.float32).eps
+        v_min_rb_ratio = v_min[0] / (v_min[2] + eps)
+        v_max_rb_ratio = v_max[0] / (v_max[2] + eps)
+        if v_min_rb_ratio < v_max_rb_ratio:
             he = np.array((v_min[:, 0], v_max[:, 0]), dtype=np.float32).T
         else:
             he = np.array((v_max[:, 0], v_min[:, 0]), dtype=np.float32).T

monai/networks/blocks/activation_checkpointing.py

@@ -0,0 +1,41 @@
+# Copyright (c) MONAI Consortium
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#     http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing import cast
+
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+
+
+class ActivationCheckpointWrapper(nn.Module):
+    """Wrapper applying activation checkpointing to a module during training.
+
+    Args:
+        module: The module to wrap with activation checkpointing.
+    """
+
+    def __init__(self, module: nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Forward pass with optional activation checkpointing.
+
+        Args:
+            x: Input tensor.
+
+        Returns:
+            Output tensor from the wrapped module.
+        """
+        return cast(torch.Tensor, checkpoint(self.module, x, use_reentrant=False))

monai/networks/nets/unet.py

@@ -17,11 +17,12 @@
 import torch
 import torch.nn as nn
 
+from monai.networks.blocks.activation_checkpointing import ActivationCheckpointWrapper
 from monai.networks.blocks.convolutions import Convolution, ResidualUnit
 from monai.networks.layers.factories import Act, Norm
 from monai.networks.layers.simplelayers import SkipConnection
 
-__all__ = ["UNet", "Unet"]
+__all__ = ["UNet", "Unet", "CheckpointUNet"]
 
 
 class UNet(nn.Module):
@@ -298,4 +299,29 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
         return x
 
 
+class CheckpointUNet(UNet):
+    """UNet variant that wraps internal connection blocks with activation checkpointing.
+
+    See `UNet` for constructor arguments. During training with gradients enabled,
+    intermediate activations inside encoder-decoder connections are recomputed in
+    the backward pass to reduce peak memory usage at the cost of extra compute.
+    """
+
+    def _get_connection_block(self, down_path: nn.Module, up_path: nn.Module, subblock: nn.Module) -> nn.Module:
+        """Returns connection block with activation checkpointing applied to all components.
+
+        Args:
+            down_path: encoding half of the layer (will be wrapped with checkpointing).
+            up_path: decoding half of the layer (will be wrapped with checkpointing).
+            subblock: block defining the next layer (will be wrapped with checkpointing).
+
+        Returns:
+            Connection block with all components wrapped for activation checkpointing.
+        """
+        subblock = ActivationCheckpointWrapper(subblock)
+        down_path = ActivationCheckpointWrapper(down_path)
+        up_path = ActivationCheckpointWrapper(up_path)
+        return super()._get_connection_block(down_path, up_path, subblock)
+
+
 Unet = UNet

monai/networks/schedulers/ddim.py

@@ -117,18 +117,14 @@ def set_timesteps(self, num_inference_steps: int, device: str | torch.device | N
             )
 
         self.num_inference_steps = num_inference_steps
-        step_ratio = self.num_train_timesteps // self.num_inference_steps
-        if self.steps_offset >= step_ratio:
-            raise ValueError(
-                f"`steps_offset`: {self.steps_offset} cannot be greater than or equal to "
-                f"`num_train_timesteps // num_inference_steps : {step_ratio}` as this will cause timesteps to exceed"
-                f" the max train timestep."
-            )
-
-        # creates integer timesteps by multiplying by ratio
-        # casting to int to avoid issues when num_inference_step is power of 3
-        timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
-        self.timesteps = torch.from_numpy(timesteps).to(device)
+        if self.steps_offset < 0 or self.steps_offset >= self.num_train_timesteps:
+            raise ValueError(f"`steps_offset`: {self.steps_offset} must be in range [0, {self.num_train_timesteps}).")
+
+        self.timesteps = (
+            torch.linspace((self.num_train_timesteps - 1) - self.steps_offset, 0, num_inference_steps, device=device)
+            .round()
+            .long()
+        )
         self.timesteps += self.steps_offset
 
     def _get_variance(self, timestep: int, prev_timestep: int) -> torch.Tensor:

monai/networks/schedulers/ddpm.py

@@ -31,7 +31,6 @@
 
 from __future__ import annotations
 
-import numpy as np
 import torch
 
 from monai.utils import StrEnum
@@ -122,11 +121,9 @@ def set_timesteps(self, num_inference_steps: int, device: str | torch.device | N
             )
 
         self.num_inference_steps = num_inference_steps
-        step_ratio = self.num_train_timesteps // self.num_inference_steps
-        # creates integer timesteps by multiplying by ratio
-        # casting to int to avoid issues when num_inference_step is power of 3
-        timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].astype(np.int64)
-        self.timesteps = torch.from_numpy(timesteps).to(device)
+        self.timesteps = (
+            torch.linspace(self.num_train_timesteps - 1, 0, self.num_inference_steps, device=device).round().long()
+        )
 
     def _get_mean(self, timestep: int, x_0: torch.Tensor, x_t: torch.Tensor) -> torch.Tensor:
         """

tests/apps/pathology/transforms/test_pathology_he_stain.py

@@ -48,7 +48,7 @@
 # input pixels not uniformly filled, leading to two different stains extracted
 EXTRACT_STAINS_TEST_CASE_5 = [
     np.array([[[100, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]]]),
-    np.array([[0.70710677, 0.18696113], [0.0, 0.0], [0.70710677, 0.98236734]]),
+    np.array([[0.18696113, 0.70710677], [0.0, 0.0], [0.98236734, 0.70710677]]),
 ]
 
 # input pixels all transparent and below the beta absorbance threshold
@@ -68,7 +68,7 @@
 NORMALIZE_STAINS_TEST_CASE_4 = [
     {"target_he": np.full((3, 2), 1)},
     np.array([[[100, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]]]),
-    np.array([[[87, 87, 87], [33, 33, 33]], [[33, 33, 33], [33, 33, 33]], [[33, 33, 33], [33, 33, 33]]]),
+    np.array([[[31, 31, 31], [85, 85, 85]], [[85, 85, 85], [85, 85, 85]], [[85, 85, 85], [85, 85, 85]]]),
 ]
 
 
@@ -135,7 +135,7 @@ def test_result_value(self, image, expected_data):
           [[0.18696113],[0],[0.98236734]] and
           [[0.70710677],[0],[0.70710677]] respectively
         - the resulting extracted stain should be
-          [[0.70710677,0.18696113],[0,0],[0.70710677,0.98236734]]
+          [[0.18696113,0.70710677],[0,0],[0.98236734,0.70710677]]
         """
         if image is None:
             with self.assertRaises(TypeError):
@@ -206,17 +206,17 @@ def test_result_value(self, arguments, image, expected_data):
 
         For test case 4:
         - For this non-uniformly filled image, the stain extracted should be
-          [[0.70710677,0.18696113],[0,0],[0.70710677,0.98236734]], as validated for the
+          [[0.18696113,0.70710677],[0,0],[0.98236734,0.70710677]], as validated for the
           ExtractHEStains class. Solving the linear least squares problem (since
           absorbance matrix = stain matrix * concentration matrix), we obtain the concentration
-          matrix that should be [[-0.3101, 7.7508, 7.7508, 7.7508, 7.7508, 7.7508],
-          [5.8022, 0, 0, 0, 0, 0]]
+          matrix that should be [[5.8022, 0, 0, 0, 0, 0],
+          [-0.3101, 7.7508, 7.7508, 7.7508, 7.7508, 7.7508]]
         - Normalizing the concentration matrix, taking the matrix product of the
           target stain matrix and the concentration matrix, using the inverse
           Beer-Lambert transform to obtain the RGB image from the absorbance
           image, and finally converting to uint8, we get that the stain normalized
-          image should be [[[87, 87, 87], [33, 33, 33]], [[33, 33, 33], [33, 33, 33]],
-          [[33, 33, 33], [33, 33, 33]]]
+          image should be [[[31, 31, 31], [85, 85, 85]], [[85, 85, 85], [85, 85, 85]],
+          [[85, 85, 85], [85, 85, 85]]]
         """
         if image is None:
             with self.assertRaises(TypeError):

tests/apps/pathology/transforms/test_pathology_he_stain_dict.py

@@ -42,7 +42,7 @@
 # input pixels not uniformly filled, leading to two different stains extracted
 EXTRACT_STAINS_TEST_CASE_5 = [
     np.array([[[100, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]]]),
-    np.array([[0.70710677, 0.18696113], [0.0, 0.0], [0.70710677, 0.98236734]]),
+    np.array([[0.18696113, 0.70710677], [0.0, 0.0], [0.98236734, 0.70710677]]),
 ]
 
 # input pixels all transparent and below the beta absorbance threshold
@@ -62,7 +62,7 @@
 NORMALIZE_STAINS_TEST_CASE_4 = [
     {"target_he": np.full((3, 2), 1)},
     np.array([[[100, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]]]),
-    np.array([[[87, 87, 87], [33, 33, 33]], [[33, 33, 33], [33, 33, 33]], [[33, 33, 33], [33, 33, 33]]]),
+    np.array([[[31, 31, 31], [85, 85, 85]], [[85, 85, 85], [85, 85, 85]], [[85, 85, 85], [85, 85, 85]]]),
 ]
 
 
@@ -129,7 +129,7 @@ def test_result_value(self, image, expected_data):
           [[0.18696113],[0],[0.98236734]] and
           [[0.70710677],[0],[0.70710677]] respectively
         - the resulting extracted stain should be
-          [[0.70710677,0.18696113],[0,0],[0.70710677,0.98236734]]
+          [[0.18696113,0.70710677],[0,0],[0.98236734,0.70710677]]
         """
         key = "image"
         if image is None:
@@ -200,17 +200,17 @@ def test_result_value(self, arguments, image, expected_data):
 
         For test case 4:
         - For this non-uniformly filled image, the stain extracted should be
-          [[0.70710677,0.18696113],[0,0],[0.70710677,0.98236734]], as validated for the
+          [[0.18696113,0.70710677],[0,0],[0.98236734,0.70710677]], as validated for the
           ExtractHEStains class. Solving the linear least squares problem (since
           absorbance matrix = stain matrix * concentration matrix), we obtain the concentration
-          matrix that should be [[-0.3101, 7.7508, 7.7508, 7.7508, 7.7508, 7.7508],
-          [5.8022, 0, 0, 0, 0, 0]]
+          matrix that should be  [[5.8022, 0, 0, 0, 0, 0],
+          [-0.3101, 7.7508, 7.7508, 7.7508, 7.7508, 7.7508]]
         - Normalizing the concentration matrix, taking the matrix product of the
           target stain matrix and the concentration matrix, using the inverse
           Beer-Lambert transform to obtain the RGB image from the absorbance
           image, and finally converting to uint8, we get that the stain normalized
-          image should be [[[87, 87, 87], [33, 33, 33]], [[33, 33, 33], [33, 33, 33]],
-          [[33, 33, 33], [33, 33, 33]]]
+          image should be [[[31, 31, 31], [85, 85, 85]], [[85, 85, 85], [85, 85, 85]],
+          [[85, 85, 85], [85, 85, 85]]]
         """
         key = "image"
         if image is None: