comodoro commited on
Commit
469e4eb
1 Parent(s): 727174c

Upload eval.py

Browse files
Files changed (1) hide show
  1. eval.py +163 -0
eval.py ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ from datasets import load_dataset, load_metric, Audio, Dataset
3
+ from transformers import pipeline, AutoFeatureExtractor
4
+ import re
5
+ import argparse
6
+ import unicodedata
7
+ from typing import Dict
8
+
9
+
10
+ def log_results(result: Dataset, args: Dict[str, str]):
11
+ """ DO NOT CHANGE. This function computes and logs the result metrics. """
12
+
13
+ log_outputs = args.log_outputs
14
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
15
+
16
+ # load metric
17
+ wer = load_metric("wer")
18
+ cer = load_metric("cer")
19
+
20
+ # compute metrics
21
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
22
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
23
+
24
+ # print & log results
25
+ result_str = (
26
+ f"WER: {wer_result}\n"
27
+ f"CER: {cer_result}"
28
+ )
29
+ print(result_str)
30
+
31
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
32
+ f.write(result_str)
33
+
34
+ # log all results in text file. Possibly interesting for analysis
35
+ if log_outputs is not None:
36
+ pred_file = f"log_{dataset_id}_predictions.txt"
37
+ target_file = f"log_{dataset_id}_targets.txt"
38
+
39
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
40
+
41
+ # mapping function to write output
42
+ def write_to_file(batch, i):
43
+ p.write(f"{i}" + "\n")
44
+ p.write(batch["prediction"] + "\n")
45
+ t.write(f"{i}" + "\n")
46
+ t.write(batch["target"] + "\n")
47
+
48
+ result.map(write_to_file, with_indices=True)
49
+
50
+
51
+ def normalize_text(text: str) -> str:
52
+ """ DO ADAPT FOR YOUR USE CASE. this function normalizes the target text. """
53
+
54
+
55
+ CHARS = {
56
+ 'ü': 'ue',
57
+ 'ö': 'oe',
58
+ 'ï': 'i',
59
+ 'ë': 'e',
60
+ 'ä': 'ae',
61
+ 'ã': 'a',
62
+ 'à': 'á',
63
+ 'ø': 'o',
64
+ 'è': 'é',
65
+ 'ê': 'é',
66
+ 'å': 'ó',
67
+ 'î': 'i',
68
+ 'ñ': 'ň',
69
+ 'ç': 's',
70
+ 'ľ': 'l',
71
+ 'ż': 'ž',
72
+ 'ł': 'w',
73
+ 'ć': 'č',
74
+ 'þ': 't',
75
+ 'ß': 'ss',
76
+ 'ę': 'en',
77
+ 'ą': 'an',
78
+ 'æ': 'ae',
79
+ }
80
+
81
+ def replace_chars(sentence):
82
+ result = ''
83
+ for ch in sentence:
84
+ new = CHARS[ch] if ch in CHARS else ch
85
+ result += new
86
+
87
+ return result
88
+
89
+ chars_to_ignore_regex = '[\,\?\.\!\-\;\:\/\"\“\„\%\”\�\–\'\`\«\»\—\’\…]'
90
+
91
+ text = text.lower()
92
+ # normalize non-standard (stylized) unicode characters
93
+ text = unicodedata.normalize('NFKC', text)
94
+ # remove punctuation
95
+ text = re.sub(chars_to_ignore_regex, "", text)
96
+ text = replace_chars(text)
97
+
98
+ # Let's also make sure we split on all kinds of newlines, spaces, etc...
99
+ text = " ".join(text.split())
100
+
101
+ return text
102
+
103
+
104
+ def main(args):
105
+ # load dataset
106
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
107
+
108
+ # for testing: only process the first two examples as a test
109
+ # dataset = dataset.select(range(10))
110
+
111
+ # load processor
112
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
113
+ sampling_rate = feature_extractor.sampling_rate
114
+
115
+ # resample audio
116
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
117
+
118
+ # load eval pipeline
119
+ asr = pipeline("automatic-speech-recognition", model=args.model_id)
120
+
121
+ # map function to decode audio
122
+ def map_to_pred(batch):
123
+ prediction = asr(batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s)
124
+
125
+ batch["prediction"] = prediction["text"]
126
+ batch["target"] = normalize_text(batch["sentence"])
127
+ return batch
128
+
129
+ # run inference on all examples
130
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
131
+
132
+ # compute and log_results
133
+ # do not change function below
134
+ log_results(result, args)
135
+
136
+
137
+ if __name__ == "__main__":
138
+ parser = argparse.ArgumentParser()
139
+
140
+ parser.add_argument(
141
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
142
+ )
143
+ parser.add_argument(
144
+ "--dataset", type=str, required=True, help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets"
145
+ )
146
+ parser.add_argument(
147
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
148
+ )
149
+ parser.add_argument(
150
+ "--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`"
151
+ )
152
+ parser.add_argument(
153
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to None. For long audio files a good value would be 5.0 seconds."
154
+ )
155
+ parser.add_argument(
156
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to None. For long audio files a good value would be 1.0 seconds."
157
+ )
158
+ parser.add_argument(
159
+ "--log_outputs", action='store_true', help="If defined, write outputs to log file for analysis."
160
+ )
161
+ args = parser.parse_args()
162
+
163
+ main(args)