Fix doc strings and minor argument update for cross entropy losses (f…

…acebookresearch#651)

Summary:
Pull Request resolved: facebookresearch#651

Updated doc strings and changed `normalized_targets` to a bool

Reviewed By: lauragustafson

Differential Revision: D24902068

fbshipit-source-id: f0031a9814dea9b73a809e111cb75f527ed8495d

classy_vision/losses/label_smoothing_loss.py

@@ -16,22 +16,25 @@
 
 @register_loss("label_smoothing_cross_entropy")
 class LabelSmoothingCrossEntropyLoss(ClassyLoss):
-    def __init__(self, ignore_index, reduction, smoothing_param):
+    def __init__(self, ignore_index=-100, reduction="mean", smoothing_param=None):
         """Intializer for the label smoothed cross entropy loss.
         This decreases gap between output scores and encourages generalization.
-        Targets provided to forward can be one-hot vectors (NxC) or class indices(Nx1)
+        Targets provided to forward can be one-hot vectors (NxC) or class indices (Nx1).
 
-        Config params:
-        'weight': weight of sample (not yet implemented),
-        'ignore_index': sample should be ignored for loss (optional),
-        'smoothing_param': value to be added to each target entry
+        This normalizes the targets to a sum of 1 based on the total count of positive
+        targets for a given sample before applying label smoothing.
+
+        Args:
+            ignore_index: sample should be ignored for loss if the class is this value
+            reduction: specifies reduction to apply to the output
+            smoothing_param: value to be added to each target entry
         """
         super().__init__()
         self._ignore_index = ignore_index
         self._reduction = reduction
         self._smoothing_param = smoothing_param
         self.loss_function = SoftTargetCrossEntropyLoss(
-            self._ignore_index, self._reduction, None
+            self._ignore_index, self._reduction, normalize_targets=False
         )
         self._eps = np.finfo(np.float32).eps
 
@@ -47,7 +50,6 @@ def from_config(cls, config: Dict[str, Any]) -> "LabelSmoothingCrossEntropyLoss"
             A LabelSmoothingCrossEntropyLoss instance.
         """
 
-        assert "weight" not in config, '"weight" not implemented'
         assert (
             "smoothing_param" in config
         ), "Label Smoothing needs a smoothing parameter"
@@ -79,7 +81,6 @@ def compute_valid_targets(self, target, classes):
         return valid_targets
 
     def smooth_targets(self, valid_targets, classes):
-
         """
         This function takes valid (No ignore values present) one-hot target vectors
         and computes smoothed target vectors (normalized) according to the loss's

classy_vision/losses/soft_target_cross_entropy_loss.py

@@ -15,19 +15,20 @@
 
 @register_loss("soft_target_cross_entropy")
 class SoftTargetCrossEntropyLoss(ClassyLoss):
-    def __init__(self, ignore_index, reduction, normalize_targets):
+    def __init__(self, ignore_index=-100, reduction="mean", normalize_targets=True):
         """Intializer for the soft target cross-entropy loss loss.
         This allows the targets for the cross entropy loss to be multilabel
 
-        Config params:
-        'weight': weight of sample (not yet implemented),
-        'ignore_index': sample should be ignored for loss (optional),
-        'reduction': specifies reduction to apply to the output (optional),
+        Args:
+            ignore_index: sample should be ignored for loss if the class is this value
+            reduction: specifies reduction to apply to the output
+            normalize_targets: whether the targets should be normalized to a sum of 1
+                based on the total count of positive targets for a given sample
         """
         super(SoftTargetCrossEntropyLoss, self).__init__()
         self._ignore_index = ignore_index
         self._reduction = reduction
-        assert normalize_targets in [None, "count_based"]
+        assert isinstance(normalize_targets, bool)
         self._normalize_targets = normalize_targets
         if self._reduction != "mean":
             raise NotImplementedError(
@@ -47,12 +48,10 @@ def from_config(cls, config: Dict[str, Any]) -> "SoftTargetCrossEntropyLoss":
             A SoftTargetCrossEntropyLoss instance.
         """
 
-        if "weight" in config:
-            raise NotImplementedError('"weight" not implemented')
         return cls(
             ignore_index=config.get("ignore_index", -100),
             reduction=config.get("reduction", "mean"),
-            normalize_targets=config.get("normalize_targets", "count_based"),
+            normalize_targets=config.get("normalize_targets", True),
         )
 
     def forward(self, output, target):
@@ -77,7 +76,7 @@ def forward(self, output, target):
         )
         valid_mask = target != self._ignore_index
         valid_targets = target.float() * valid_mask.float()
-        if self._normalize_targets == "count_based":
+        if self._normalize_targets:
             valid_targets /= self._eps + valid_targets.sum(dim=1, keepdim=True)
         per_sample_per_target_loss = -valid_targets * F.log_softmax(output, -1)
         # perform reduction

test/losses_soft_target_cross_entropy_loss_test.py

@@ -59,7 +59,7 @@ def test_unnormalized_soft_target_cross_entropy(self):
             "name": "soft_target_cross_entropy",
             "ignore_index": -1,
             "reduction": "mean",
-            "normalize_targets": None,
+            "normalize_targets": False,
         }
         crit = SoftTargetCrossEntropyLoss.from_config(config)
         outputs = self._get_outputs()