2 SinusoidalPositionalEmbedding fixes (#8226)

src/transformers/modeling_bart.py

@@ -1328,8 +1328,6 @@ class SinusoidalPositionalEmbedding(nn.Embedding):
 
     def __init__(self, num_positions, embedding_dim, padding_idx=None):
         super().__init__(num_positions, embedding_dim)
-        if embedding_dim % 2 != 0:
-            raise NotImplementedError(f"odd embedding_dim {embedding_dim} not supported")
         self.weight = self._init_weight(self.weight)
 
     @staticmethod
@@ -1342,10 +1340,11 @@ def _init_weight(out: nn.Parameter):
         position_enc = np.array(
             [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
         )
-        out[:, 0 : dim // 2] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))  # This line breaks for odd n_pos
-        out[:, dim // 2 :] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
+        out.requires_grad = False  # set early to avoid an error in pytorch-1.8+
+        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
+        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
+        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
         out.detach_()
-        out.requires_grad = False
         return out
 
     @torch.no_grad()

tests/test_modeling_bart.py

@@ -620,8 +620,8 @@ def test_positional_emb_cache_logic(self):
         self.assertListEqual(no_cache[-1].tolist(), yes_cache[0][0].tolist())
 
     def test_odd_embed_dim(self):
-        with self.assertRaises(NotImplementedError):
-            SinusoidalPositionalEmbedding(num_positions=4, embedding_dim=5, padding_idx=0).to(torch_device)
+        # odd embedding_dim is allowed
+        SinusoidalPositionalEmbedding(num_positions=4, embedding_dim=5, padding_idx=0).to(torch_device)
 
         # odd num_positions is allowed
         SinusoidalPositionalEmbedding(num_positions=5, embedding_dim=4, padding_idx=0).to(torch_device)