Add classes/ignore_index to losses (#11)

Author: isaaccorley

tests/test_losses.py

@@ -5,6 +5,7 @@
 import torchseg.losses._functional as F
 from torchseg.losses import (
     DiceLoss,
+    FocalLoss,
     JaccardLoss,
     MCCLoss,
     SoftBCEWithLogitsLoss,
@@ -333,3 +334,15 @@ def test_binary_mcc_loss():
 
     loss = criterion(y_pred, y_true)
     assert float(loss) == pytest.approx(0.5, abs=eps)
+
+
+@torch.no_grad()
+@pytest.mark.parametrize("loss_fn", [DiceLoss, JaccardLoss, FocalLoss])
+@pytest.mark.parametrize("classes", [None, [1]])
+@pytest.mark.parametrize("ignore_index", [None, 0, -255])
+def test_classes_arg(loss_fn, classes, ignore_index):
+    criterion = loss_fn(mode="multiclass", classes=classes, ignore_index=ignore_index)
+    y_pred = torch.zeros(1, 2, 128, 128, dtype=torch.float)
+    y_pred[:, 0, ...] = 1.0
+    y_true = torch.ones(1, 128, 128, dtype=torch.long)
+    criterion(y_pred, y_true)

torchseg/losses/focal.py

@@ -18,6 +18,7 @@ def __init__(
         reduction: Optional[str] = "mean",
         normalized: bool = False,
         reduced_threshold: Optional[float] = None,
+        classes: Optional[list[int]] = None,
     ):
         """Compute Focal loss
 
@@ -30,6 +31,8 @@ def __init__(
             normalized: Use normalized focal loss (https://arxiv.org/pdf/1909.07829.pdf)
             reduced_threshold: Switch to reduced focal loss.
                 Note, when using this mode you should use `reduction="sum"`.
+            classes:  List of classes that contribute in loss computation.
+                By default, all channels are included. Only supported in multiclass mode
 
         Shape
              - **y_pred** - torch.Tensor of shape (N, C, H, W)
@@ -44,6 +47,7 @@ def __init__(
 
         self.mode = mode
         self.ignore_index = ignore_index
+        self.classes = classes
         self.focal_loss_fn = partial(
             focal_loss_with_logits,
             alpha=alpha,
@@ -75,13 +79,14 @@ def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
                 not_ignored = y_true != self.ignore_index
 
             for cls in range(num_classes):
-                cls_y_true = (y_true == cls).long()
-                cls_y_pred = y_pred[:, cls, ...]
+                if self.classes is None or cls in self.classes:
+                    cls_y_true = (y_true == cls).long()
+                    cls_y_pred = y_pred[:, cls, ...]
 
-                if self.ignore_index is not None:
-                    cls_y_true = cls_y_true[not_ignored]
-                    cls_y_pred = cls_y_pred[not_ignored]
+                    if self.ignore_index is not None:
+                        cls_y_true = cls_y_true[not_ignored]
+                        cls_y_pred = cls_y_pred[not_ignored]
 
-                loss += self.focal_loss_fn(cls_y_pred, cls_y_true)
+                    loss += self.focal_loss_fn(cls_y_pred, cls_y_true)
 
         return loss

torchseg/losses/jaccard.py

@@ -16,6 +16,7 @@ def __init__(
         log_loss: bool = False,
         from_logits: bool = True,
         smooth: float = 0.0,
+        ignore_index: Optional[int] = None,
         eps: float = 1e-7,
     ):
         """Jaccard loss for image segmentation task.
@@ -29,6 +30,8 @@ def __init__(
                 otherwise `1 - jaccard_coeff`
             from_logits: If True, assumes input is raw logits
             smooth: Smoothness constant for dice coefficient
+            ignore_index: Label that indicates ignored pixels
+                (does not contribute to loss)
             eps: A small epsilon for numerical stability to avoid zero division error
                 (denominator will be always greater or equal to eps)
 
@@ -53,6 +56,7 @@ def __init__(
         self.from_logits = from_logits
         self.smooth = smooth
         self.eps = eps
+        self.ignore_index = ignore_index
         self.log_loss = log_loss
 
     def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
@@ -76,17 +80,36 @@ def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
             y_true = y_true.view(bs, 1, -1)
             y_pred = y_pred.view(bs, 1, -1)
 
+            if self.ignore_index is not None:
+                mask = y_true != self.ignore_index
+                y_pred = y_pred * mask
+                y_true = y_true * mask
+
         if self.mode == MULTICLASS_MODE:
             y_true = y_true.view(bs, -1)
             y_pred = y_pred.view(bs, num_classes, -1)
 
-            y_true = F.one_hot(y_true, num_classes)  # N,H*W -> N,H*W, C
-            y_true = y_true.permute(0, 2, 1)  # H, C, H*W
+            if self.ignore_index is not None:
+                mask = y_true != self.ignore_index
+                y_pred = y_pred * mask.unsqueeze(1)
+
+                y_true = F.one_hot(
+                    (y_true * mask).to(torch.long), num_classes
+                )  # N,H*W -> N,H*W, C
+                y_true = y_true.permute(0, 2, 1) * mask.unsqueeze(1)  # N, C, H*W
+            else:
+                y_true = F.one_hot(y_true, num_classes)  # N,H*W -> N,H*W, C
+                y_true = y_true.permute(0, 2, 1)  # H, C, H*W
 
         if self.mode == MULTILABEL_MODE:
             y_true = y_true.view(bs, num_classes, -1)
             y_pred = y_pred.view(bs, num_classes, -1)
 
+            if self.ignore_index is not None:
+                mask = y_true != self.ignore_index
+                y_pred = y_pred * mask
+                y_true = y_true * mask
+
         scores = soft_jaccard_score(
             y_pred,
             y_true.type(y_pred.dtype),