Update eval.py

eval.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 import argparse
-import os
 import re
 from typing import Dict
 
@@ -12,6 +11,8 @@ from slugify import slugify
 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."""
@@ -19,7 +20,11 @@ def log_results(result: Dataset, args: Dict[str, str]):
     log_outputs = args.log_outputs
     lm = "withLM" if args.use_lm else "noLM"
     model_id = args.model_id.replace("/", "_").replace(".", "")
-    dataset_id = "_".join([model_id] + args.dataset.split("/") + [args.config, slugify(args.filter), args.split, lm])
+    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")
@@ -58,24 +63,36 @@ def normalize_text(original_text: str, dataset: str) -> str:
     text = original_text.lower()
     if dataset.lower().endswith("fleurs"): 
         replacements = (
-            ("e.kr", "etter kristus fødsel"),
-            ("f.kr", "før kristi fødsel"),
-            ("km/t", "kilometer i timen"),
-            ("km", "kilometer"),
-            ("cm", "centimeter"),
-            ("mm", "millimeter"),
-            ("kl.", "klokka"),
-            ("f.eks", "for eksempel"),
+            (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(f' {abrev}', f" {expasion}", text)
-        text = re.sub(':00', '', text)
-        text = re.sub(r"(\d{1,2})[ .](\d{3})", r"\1\2", text)
-        text = re.sub(r"(\d{2}):(\d{2})", r"\1 \2", text)
-        text = re.compile(r"-?[1-9][\d.]*").sub(lambda x: n2w(x.group(0), lang="no"), text.replace(".", ""))
-
-    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) + " "
+            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(".", ""))
+
+
+    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()
@@ -109,6 +126,7 @@ def normalize_text(original_text: str, dataset: str) -> str:
         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('<ee>', 'eee', text)
     text = re.sub('<qq>', 'qqq', text)
@@ -131,8 +149,7 @@ def main(args):
     if args.filter:
         attribute, value = list(map(str.strip, args.filter.split(":")))
         dataset = dataset.filter(
-            lambda x: x[attribute] == value,
-            num_proc=os.cpu_count() // 2,
+            lambda x: x[attribute == value],
             desc=f"Filtering on {args.filter}",
         )
     # for testing: only process the first two examples as a test
@@ -233,4 +250,3 @@ if __name__ == "__main__":
     args = parser.parse_args()
 
     main(args)
-