Update eval.py

eval.py

@@ -5,16 +5,26 @@ from typing import Dict
 
 import torch
 from datasets import Audio, Dataset, load_dataset, load_metric
+from num2words import num2words as n2w
+from slugify import slugify
 
-from transformers import AutoFeatureExtractor, pipeline
+from transformers import AutoFeatureExtractor, AutoModelForCTC, pipeline, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM, Wav2Vec2FeatureExtractor
+# from pyctcdecode import BeamSearchDecoderCTC
+
+from cardinal_numbers import convert_nums
 
 
 def log_results(result: Dataset, args: Dict[str, str]):
     """DO NOT CHANGE. This function computes and logs the result metrics."""
 
     log_outputs = args.log_outputs
+    lm = "withLM" if args.use_lm else "noLM"
     model_id = args.model_id.replace("/", "_").replace(".", "")
-    dataset_id = "_".join(args.dataset.split("/") + [model_id, args.config, args.split])
+    if args.filter:
+        extra_args = [args.config, slugify(args.filter), args.split, lm]
+    else:
+        extra_args = [args.config, args.split, lm]
+    dataset_id = "_".join([model_id] + args.dataset.split("/") + extra_args)
 
     # load metric
     wer = load_metric("wer")
@@ -30,6 +40,8 @@ def log_results(result: Dataset, args: Dict[str, str]):
 
     with open(f"{dataset_id}_eval_results.txt", "w") as f:
         f.write(result_str)
+    with open(f"{dataset_id}_eval_results.tsv", "w") as f:
+        f.write("\t".join([args.model_id, args.dataset, args.config, args.filter, args.split, str(lm), str(wer_result), str(cer_result)]))
 
     # log all results in text file. Possibly interesting for analysis
     if log_outputs is not None:
@@ -37,7 +49,6 @@ def log_results(result: Dataset, args: Dict[str, str]):
         target_file = f"log_{dataset_id}_targets.txt"
 
         with open(pred_file, "w") as p, open(target_file, "w") as t:
-
             # mapping function to write output
             def write_to_file(batch, i):
                 p.write(f"{i}" + "\n")
@@ -48,24 +59,80 @@ def log_results(result: Dataset, args: Dict[str, str]):
             result.map(write_to_file, with_indices=True)
 
 
-def normalize_text(text: str) -> str:
+def normalize_text(original_text: str, dataset: str) -> str:
     """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
 
-    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+    text = original_text.lower()
+    if dataset.lower().endswith("fleurs"): 
+        replacements = (
+            (r"\be\.kr", "etter kristus fødsel"),
+            (r"\bf\.kr", "før kristi fødsel"),
+            (r"\bca[.]?\b", "circa"),
+            (r"(\d)\s*km/t", r"\1 kilometer i timen"),
+            (r"(\d)\s*km", r"\1 kilometer"),
+            (r"(\d)\s*cm", r"\1 centimeter"),
+            (r"(\d)\s*mm", r"\1 millimeter"),
+            (r"kl\.", "klokka"),
+            (r"f\.eks", "for eksempel"),
+        )
+        for abrev, expasion in replacements:
+            text = re.sub(abrev, expasion, text)
+        text = re.sub(r'(\d+)[-–](\d+)', r'\1 til \2', text)  # 1-89, 70-90
+        text = re.sub(r'(\d{2}):00', r'\1', text)  # 21:00
+        text = re.sub(r"(\d{2}):0(\d{1})", r"\1 null \2", text)  # 17:03
+        text = re.sub(r"(\d{1,2}):(\d{1,2})", r"\1 \2", text)  # 17:23 (time), 4:3 (aspect ratios)
+        text = re.sub(r"(1[1-9])00", r"\1 hundre", text)  # 1800, 1900
+        text = re.sub(r"(1[1-9])0([1-9])", r"\1 null \2 ", text)  # 1901, 1909
+        text = re.sub(r"(1[1-9])([1-9]\d)", r"\1 \2 ", text)  # 1911, 1987
+        text = re.sub(r"(20)0([1-9])", r"\1 null \2 ", text)  # 2009
+        text = re.sub(r"(20)(\d{2})", r"\1 \2 ", text)  # 2009
+        text = re.sub(r"(\d{1,3})[.](\d{1,2})", r"\1 dot \2 ", text)  # 802.11n, 2.5ghz (in English)
+        text = re.sub(r"(\d{1,2})[ .](\d{3})", r"\1\2", text)  # 10 000, 32.000
+        text = re.sub(r'(\w+)-(\w+)', r'\1 \2', text)  # n-standard
+        # text = re.compile(r"-?0?[1-9][\d.]*").sub(lambda x: n2w(x.group(0), lang="no"), text.replace(".", ""))
+        text = re.compile(r"-?0?[1-9][\d.]*").sub(lambda x: convert_nums(int(x.group(0)), nn=True), text.replace(".", ""))
+
 
-    text = re.sub(chars_to_ignore_regex, "", text.lower()) + " "
-    text = re.sub('[áàâ]', 'a', text)
-    text = re.sub('[ä]', 'æ', text)
-    text = re.sub('[éèëê]', 'e', text)
-    text = re.sub('[íìïî]', 'i', text)
-    text = re.sub('[óòöô]', 'o', text)
-    text = re.sub('[ö]', 'ø', text)
-    text = re.sub('[ç]', 'c', text)
-    text = re.sub('[úùüû]', 'u', text)
-    text = re.sub('\s', ' ', text)
-    text = re.sub('<ee>', 'eee', text)
-    text = re.sub('<qq>', 'qqq', text)
-    text = re.sub('<mm>', 'mmm', text)
+    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+    text = re.sub(chars_to_ignore_regex, "", text) + " "
+
+    if dataset.lower().endswith("nst"):
+        text = text.lower()
+        text = text.replace("(...vær stille under dette opptaket...)", "")
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
+        text = re.sub('\s+', ' ', text)
+    elif dataset.lower().endswith("npsc"):
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        text = re.sub('\s+', ' ', text)
+    elif dataset.lower().endswith("fleurs"): 
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        text = re.sub('[«»]', '', text)
+        text = re.sub('\s+', ' ', text)
+    text = re.sub('<e+h?>', 'eee', text)
+    text = re.sub('<m+>', 'mmm', text)
+    text = re.sub('<q+>', 'qqq', text)
     text = re.sub('<inaudible>', 'xxx', text)
     text = re.sub('[<>]', '', text)
 
@@ -76,13 +143,18 @@ def normalize_text(text: str) -> str:
     # for t in token_sequences_to_ignore:
     #     text = " ".join(text.split(t))
 
-    return text
+    return text.strip()
 
 
 def main(args):
     # load dataset
     dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
-
+    if args.filter:
+        attribute, value = list(map(str.strip, args.filter.split(":")))
+        dataset = dataset.filter(
+            lambda x: x[attribute] == value,
+            desc=f"Filtering on {args.filter}",
+        )
     # for testing: only process the first two examples as a test
     # dataset = dataset.select(range(10))
 
@@ -96,7 +168,29 @@ def main(args):
     # load eval pipeline
     if args.device is None:
         args.device = 0 if torch.cuda.is_available() else -1
-    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+    # asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    model_instance = AutoModelForCTC.from_pretrained(args.model_id)
+    if args.use_lm:
+        processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
+        decoder = processor.decoder
+    else:
+        processor = Wav2Vec2Processor.from_pretrained(args.model_id)
+        decoder = None
+    asr = pipeline(
+        "automatic-speech-recognition",
+        model=model_instance,
+        tokenizer=processor.tokenizer, 
+        feature_extractor=processor.feature_extractor,
+        decoder=decoder,
+        device=args.device
+    )
+
+    # feature_extractor_dict, _ = Wav2Vec2FeatureExtractor.get_feature_extractor_dict(args.model_id)
+    # feature_extractor_dict["processor_class"] = "Wav2Vec2Processor" if not args.use_lm else "Wav2Vec2ProcessorWithLM"
+    # feature_extractor = Wav2Vec2FeatureExtractor.from_dict(feature_extractor_dict)
+
+    # asr = pipeline("automatic-speech-recognition", model=args.model_id, feature_extractor=feature_extractor, device=args.device, decoder=BeamSearchDecoderCTC.load_from_dir("./"))
 
     # map function to decode audio
     def map_to_pred(batch):
@@ -105,7 +199,7 @@ def main(args):
         )
 
         batch["prediction"] = prediction["text"]
-        batch["target"] = normalize_text(batch["text"])
+        batch["target"] = normalize_text(batch[args.text_column], args.dataset)
         return batch
 
     # run inference on all examples
@@ -131,7 +225,13 @@ if __name__ == "__main__":
     parser.add_argument(
         "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
     )
+    parser.add_argument(
+        "--filter", type=str, default="", help="Simple filter on attributes. *E.g.* `region_of_youth:Troms` would pnly keep those samplesfor which the condition is met"
+    )
     parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--text_column", type=str, default="text", help="Column name containing the transcription."
+    )
     parser.add_argument(
         "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
     )
@@ -147,6 +247,9 @@ if __name__ == "__main__":
         default=None,
         help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
     )
+    parser.add_argument(
+        "--use_lm", action="store_true", help="If defined, use included language model as the decoder."
+    )
     args = parser.parse_args()
 
     main(args)