The neural network initialization bug is fixed.

modeling_hierarchical_classifier.py

@@ -77,7 +77,7 @@ class DistanceBasedLogisticLoss(_Loss):
         inputs = inputs.view(-1)
         targets = targets.to(inputs.dtype).view(-1)
         p = distance_to_probability(inputs, self.margin)
-        return 1.0 - torch.nn.functional.binary_cross_entropy(input=p, target=targets, reduction=self.reduction)
+        return torch.nn.functional.binary_cross_entropy(input=p, target=targets, reduction=self.reduction)
 
 
 class LayerGatingNetwork(torch.nn.Module):
@@ -128,11 +128,6 @@ class XLMRobertaXLForHierarchicalEmbedding(XLMRobertaXLPreTrainedModel, ABC):
 
         self.init_weights()
 
-    def init_weights(self):
-        super().init_weights()
-        with torch.no_grad():
-            self.layer_weights.reset_parameters()
-
     def forward(
             self,
             input_ids: Optional[torch.LongTensor] = None,