Add support for <=PyTorch1.9.1 (#5467)

Fixes #5469  

### Description

For CrossEntropy, PyTorch v1.10+ allows target values to be either class
probabilities or indices, whereas v1.9- only allows for class indices,
so class indices are now used. This should provide a better solution for
#5393 and remove the need for #5401


### Types of changes
<!--- Put an `x` in all the boxes that apply, and remove the not
applicable items -->
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).

Signed-off-by: JHancox <48477639+JHancox@users.noreply.github.com>
Signed-off-by: Behrooz <3968947+drbeh@users.noreply.github.com>
Co-authored-by: Behrooz <3968947+drbeh@users.noreply.github.com>

Author: 2 people

monai/apps/pathology/losses/hovernet_loss.py

@@ -121,7 +121,9 @@ def forward(self, prediction: Dict[str, torch.Tensor], target: Dict[str, torch.T
         dice_loss_np = (
             self.dice(prediction[HoVerNetBranch.NP.value], target[HoVerNetBranch.NP.value]) * self.lambda_np_dice
         )
-        ce_loss_np = self.ce(prediction[HoVerNetBranch.NP.value], target[HoVerNetBranch.NP.value]) * self.lambda_np_ce
+        # convert to target class indices
+        argmax_target = target[HoVerNetBranch.NP.value].argmax(dim=1)
+        ce_loss_np = self.ce(prediction[HoVerNetBranch.NP.value], argmax_target) * self.lambda_np_ce
         loss_np = dice_loss_np + ce_loss_np
 
         # Compute the HV branch loss
@@ -146,9 +148,9 @@ def forward(self, prediction: Dict[str, torch.Tensor], target: Dict[str, torch.T
             dice_loss_nc = (
                 self.dice(prediction[HoVerNetBranch.NC.value], target[HoVerNetBranch.NC.value]) * self.lambda_nc_dice
             )
-            ce_loss_nc = (
-                self.ce(prediction[HoVerNetBranch.NC.value], target[HoVerNetBranch.NC.value]) * self.lambda_nc_ce
-            )
+            # Convert to target class indices
+            argmax_target = target[HoVerNetBranch.NC.value].argmax(dim=1)
+            ce_loss_nc = self.ce(prediction[HoVerNetBranch.NC.value], argmax_target) * self.lambda_nc_ce
             loss_nc = dice_loss_nc + ce_loss_nc
 
         # Sum the losses from each branch

tests/test_hovernet_loss.py

@@ -21,7 +21,6 @@
 from monai.transforms import GaussianSmooth, Rotate
 from monai.transforms.intensity.array import ComputeHoVerMaps
 from monai.utils.enums import HoVerNetBranch
-from tests.utils import SkipIfBeforePyTorchVersion
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
@@ -169,7 +168,6 @@ def test_shape_generator(num_classes=1, num_objects=3, batch_size=1, height=5, w
 ]
 
 
-@SkipIfBeforePyTorchVersion((1, 10))
 class TestHoverNetLoss(unittest.TestCase):
     @parameterized.expand(CASES)
     def test_shape(self, input_param, expected_loss):