last commit for the day

hl_gauss_pytorch/hl_gauss.py

@@ -73,16 +73,34 @@ def transform_to_probs(self, target):
     def transform_from_probs(self, probs):
         return (probs * self.centers).sum(dim = -1)
 
+    def forward_kl_div(
+        self,
+        pred,
+        target
+    ):
+        """
+        allow for predicted value to be passed in, in which it will also be binned to probs and kl div used with target
+        """
+        assert pred.shape == target.shape
+
+        logprob_pred = self.transform_to_logprobs(pred)
+        logprob_target = self.transform_to_logprobs(target)
+
+        return F.kl_div(logprob_pred, logprob_target, log_target = True, reduction = 'mean')
+
     @torch.autocast('cuda', enabled = False)
     def forward(
         self,
         logits,
         target = None
     ):
-        assert logits.shape[-1] == self.num_bins
-
         return_loss = exists(target)
 
+        if return_loss and logits.shape == target.shape:
+            return self.forward_kl_div(logits, target)
+
+        assert logits.shape[-1] == self.num_bins
+
         if return_loss:
             target_probs = self.transform_to_probs(target)
             return F.cross_entropy(logits, target_probs)

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hl-gauss-pytorch"
-version = "0.1.4"
+version = "0.1.5"
 description = "HL Gauss - Pytorch"
 authors = [
     { name = "Phil Wang", email = "[email protected]" }