Use processor.tokenizer to get labels

scripts/helpers/asr.py

@@ -147,25 +147,18 @@ def get_metrics_computer(processor):
     def compute_metrics(pred):
 
         pred_logits = pred.predictions
-        pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id
 
         if type(processor).__name__ == "Wav2Vec2ProcessorWithLM":
-
             pred_str    = processor.batch_decode(pred_logits).text
-
-            label_chars = [ processor.tokenizer.convert_ids_to_tokens(l) for l in pred.label_ids ]
-            label_str   = [ "".join([ id for id in l if id not in processor.tokenizer.unique_no_split_tokens ]) for l in label_chars ]
-            label_str   = [ l.replace(processor.tokenizer.word_delimiter_token, " ").strip() for l in label_str ]
-
         else:
-
-            pred_logits = pred.predictions
             pred_ids = np.argmax(pred_logits, axis=-1)
-
             pred_str = processor.batch_decode(pred_ids)
-            # we do not want to group tokens when computing the metrics
-            label_str = processor.batch_decode(pred.label_ids, group_tokens=False)
 
+        # Replace data collator padding with tokenizer's padding
+        pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id
+        # Retrieve labels as characters, e.g. 'hello', from label_ids, e.g. [5, 3, 10, 10, 2] (where 5 = 'h')
+        label_str = processor.tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
         wer = wer_metric.compute(predictions=pred_str, references=label_str)
 
         return {"wer": wer}

scripts/train_asr-by-w2v2-ft.py

@@ -6,7 +6,7 @@
 from datasets import load_metric
 from helpers.asr import (
     configure_lm,
-    configure_w2v2_components,
+    configure_w2v2_for_training,
     DataCollatorCTCWithPadding,
     dataset_from_dict,
     get_metrics_computer,
@@ -66,7 +66,7 @@
 
 dataset, vocab_dict = preprocess_text(dataset)
 
-model, processor = configure_w2v2_components(dataset, args, vocab_dict, w2v2_config)
+model, processor = configure_w2v2_for_training(dataset, args, vocab_dict, w2v2_config)
 
 if args.lm_arpa is not None:
     processor = configure_lm(processor, args.lm_arpa, args.output_dir)