Merge pull request #189 from Adamits/cuda-numpy-ser-hotfix

Cuda numpy ser hotfix

yoyodyne/evaluators.py

@@ -197,15 +197,15 @@ class SEREvaluator(Evaluator):
 
     def _compute_ser(
         self,
-        preds: List[str],
-        target: List[str],
+        preds: torch.Tensor,
+        target: torch.Tensor,
     ) -> float:
         errors = self._edit_distance(preds, target)
         total = len(target)
         return errors / total
 
     @staticmethod
-    def _edit_distance(x: List[str], y: List[str]) -> int:
+    def _edit_distance(x: torch.Tensor, y: torch.Tensor) -> int:
         idim = len(x) + 1
         jdim = len(y) + 1
         table = numpy.zeros((idim, jdim), dtype=numpy.uint16)
@@ -224,8 +224,8 @@ def _edit_distance(x: List[str], y: List[str]) -> int:
 
     def get_eval_item(
         self,
-        predictions: List[List[str]],
-        golds: List[List[str]],
+        predictions: torch.Tensor,
+        golds: torch.Tensor,
         pad_idx: int,
     ) -> EvalItem:
         sers = [self._compute_ser(p, g) for p, g in zip(predictions, golds)]
@@ -236,13 +236,10 @@ def _finalize_tensor(
         tensor: torch.Tensor,
         end_idx: int,
         pad_idx: int,
-    ) -> List[List[str]]:
+    ) -> torch.Tensor:
         # Not necessary if batch size is 1.
         if tensor.size(0) == 1:
-            # Returns a list of a numpy char vector.
-            # This is allows evaluation over strings without converting
-            # integer indices back to symbols.
-            return [numpy.char.mod("%d", tensor.cpu().numpy())]
+            return [tensor]
         out = []
         for prediction in tensor:
             # Gets first instance of EOS.
@@ -253,23 +250,22 @@ def _finalize_tensor(
                 eos = eos[0]
             else:
                 # Leaves tensor[i] alone.
-                out.append(numpy.char.mod("%d", prediction))
+                out.append(prediction)
                 continue
             # Hack in case the first prediction is EOS. In this case
             # torch.split will result in an error, so we change these 0's to
             # 1's, which will make the entire sequence EOS as intended.
             eos[eos == 0] = 1
             symbols, *_ = torch.split(prediction, eos)
-            # Accumulates a list of numpy char vectors.
-            out.append(numpy.char.mod("%d", symbols))
+            out.append(symbols)
         return out
 
     def finalize_predictions(
         self,
         predictions: torch.Tensor,
         end_idx: int,
         pad_idx: int,
-    ) -> List[List[str]]:
+    ) -> torch.Tensor:
         """Finalizes predictions.
 
         Args:
@@ -287,7 +283,7 @@ def finalize_golds(
         golds: torch.Tensor,
         end_idx: int,
         pad_idx: int,
-    ):
+    ) -> torch.Tensor:
         return self._finalize_tensor(golds, end_idx, pad_idx)
 
     @property