Fix broken bleu metrics (#228)

lightning_transformers/task/nlp/translation/model.py

@@ -42,10 +42,7 @@ def compute_generate_metrics(self, batch, prefix):
         tgt_lns = self.tokenize_labels(batch["labels"])
         pred_lns = self.generate(batch["input_ids"], batch["attention_mask"])
         # wrap targets in list as score expects a list of potential references
-        tgt_tokens = tuple([tuple(tuple(reference.split()) for reference in tgt_lns)])
-        pred_tokens = tuple(tuple(prediction.split()) for prediction in pred_lns)
-        result = self.bleu(tgt_tokens, pred_tokens)
-
+        result = self.bleu(pred_lns, tgt_lns)
         self.log(f"{prefix}_bleu_score", result, on_step=False, on_epoch=True, prog_bar=True)
 
     def configure_metrics(self, stage: str):

requirements.txt

@@ -4,7 +4,7 @@ numpy
 tqdm
 
 #metrics
-torchmetrics>=0.5.1
+torchmetrics>=0.7.0
 
 # huggingface
 transformers

requirements/docs.txt

@@ -11,3 +11,4 @@ sphinx-autodoc-typehints>=1.11.1  # required by: https://github.com/agronholm/sp
 sphinx-paramlinks>=0.4.0
 sphinx-togglebutton>=0.2
 sphinx-copybutton>=0.3
+mistune==0.8.4

tests/core/test_loggers.py

@@ -16,7 +16,8 @@ def setUp(self):
         # run logger offline
         self.logger = WABLogger(offline=True)
 
-    def tearDown(self):
+    @classmethod
+    def tearDownClass(cls):
         # delete wandb folder
         cwd = os.getcwd()
         if os.path.exists(os.path.join(cwd, "wandb")):