fix dtype error

sent1 was float64 and sent2 float32 in some cases
mschrimpf · Feb 19, 2021 · a95bded · a95bded
1 parent b13982e
commit a95bded
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/neural_nlp/models/implementations.py b/neural_nlp/models/implementations.py
@@ -931,8 +931,8 @@ def glue_dataset(self, examples, label_list, output_mode):
             sents1 = [self._sent_mean(self._encode_sentence(sent)) for sent in tqdm(text_a)]
             sents2 = [self._sent_mean(self._encode_sentence(sent)) for sent in tqdm(text_b)]
             for sent1, sent2 in zip(sents1, sents2):
-                sent1 = torch.tensor(sent1)
-                sent2 = torch.tensor(sent2)
+                sent1 = torch.tensor(sent1, dtype=torch.float64)
+                sent2 = torch.tensor(sent2, dtype=torch.float64)
                 f = torch.cat([sent1, sent2, torch.abs(sent1 - sent2), sent1 * sent2], -1)
                 features.append(PytorchWrapper._tensor_to_numpy(f))
             all_features = torch.tensor(features).float()