versae commited on
Commit
fc623b7
1 Parent(s): 0121f5d

Test results before lang model

Browse files
NbAiLab_NPSC_16K_mp3_bokmaal_test_eval_results_no_lang_model.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ WER: 0.10098434478047091
2
+ CER: 0.03245403756734263
eval.py ADDED
@@ -0,0 +1,151 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ import torch
7
+ from datasets import Audio, Dataset, load_dataset, load_metric
8
+
9
+ from transformers import AutoFeatureExtractor, pipeline
10
+
11
+
12
+ def log_results(result: Dataset, args: Dict[str, str]):
13
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
14
+
15
+ log_outputs = args.log_outputs
16
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
17
+
18
+ # load metric
19
+ wer = load_metric("wer")
20
+ cer = load_metric("cer")
21
+
22
+ # compute metrics
23
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
24
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
25
+
26
+ # print & log results
27
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
28
+ print(result_str)
29
+
30
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
31
+ f.write(result_str)
32
+
33
+ # log all results in text file. Possibly interesting for analysis
34
+ if log_outputs is not None:
35
+ pred_file = f"log_{dataset_id}_predictions.txt"
36
+ target_file = f"log_{dataset_id}_targets.txt"
37
+
38
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
39
+
40
+ # mapping function to write output
41
+ def write_to_file(batch, i):
42
+ p.write(f"{i}" + "\n")
43
+ p.write(batch["prediction"] + "\n")
44
+ t.write(f"{i}" + "\n")
45
+ t.write(batch["target"] + "\n")
46
+
47
+ result.map(write_to_file, with_indices=True)
48
+
49
+
50
+ def normalize_text(text: str) -> str:
51
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
52
+
53
+ chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
54
+
55
+ text = re.sub(chars_to_ignore_regex, "", text.lower()) + " "
56
+ text = re.sub('[áàâ]', 'a', text)
57
+ text = re.sub('[ä]', 'æ', text)
58
+ text = re.sub('[éèëê]', 'e', text)
59
+ text = re.sub('[íìïî]', 'i', text)
60
+ text = re.sub('[óòöô]', 'o', text)
61
+ text = re.sub('[ö]', 'ø', text)
62
+ text = re.sub('[ç]', 'c', text)
63
+ text = re.sub('[úùüû]', 'u', text)
64
+ text = re.sub('\s', ' ', text)
65
+ text = re.sub('<ee>', 'eee', text)
66
+ text = re.sub('<qq>', 'qqq', text)
67
+ text = re.sub('<mm>', 'mmm', text)
68
+ text = re.sub('<inaudible>', 'xxx', text)
69
+ text = re.sub('[<>]', '', text)
70
+
71
+ # # In addition, we can normalize the target text, e.g. removing new lines characters etc...
72
+ # # note that order is important here!
73
+ # token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
74
+
75
+ # for t in token_sequences_to_ignore:
76
+ # text = " ".join(text.split(t))
77
+
78
+ return text
79
+
80
+
81
+ def main(args):
82
+ # load dataset
83
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
84
+
85
+ # for testing: only process the first two examples as a test
86
+ # dataset = dataset.select(range(10))
87
+
88
+ # load processor
89
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
90
+ sampling_rate = feature_extractor.sampling_rate
91
+
92
+ # resample audio
93
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
94
+
95
+ # load eval pipeline
96
+ if args.device is None:
97
+ args.device = 0 if torch.cuda.is_available() else -1
98
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
99
+
100
+ # map function to decode audio
101
+ def map_to_pred(batch):
102
+ prediction = asr(
103
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
104
+ )
105
+
106
+ batch["prediction"] = prediction["text"]
107
+ batch["target"] = normalize_text(batch["text"])
108
+ return batch
109
+
110
+ # run inference on all examples
111
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
112
+
113
+ # compute and log_results
114
+ # do not change function below
115
+ log_results(result, args)
116
+
117
+
118
+ if __name__ == "__main__":
119
+ parser = argparse.ArgumentParser()
120
+
121
+ parser.add_argument(
122
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
123
+ )
124
+ parser.add_argument(
125
+ "--dataset",
126
+ type=str,
127
+ required=True,
128
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
129
+ )
130
+ parser.add_argument(
131
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
132
+ )
133
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
134
+ parser.add_argument(
135
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
136
+ )
137
+ parser.add_argument(
138
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
139
+ )
140
+ parser.add_argument(
141
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
142
+ )
143
+ parser.add_argument(
144
+ "--device",
145
+ type=int,
146
+ default=None,
147
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
148
+ )
149
+ args = parser.parse_args()
150
+
151
+ main(args)
log_NbAiLab_NPSC_16K_mp3_bokmaal_test_predictions_no_lang_model.txt ADDED
The diff for this file is too large to render. See raw diff
 
log_NbAiLab_NPSC_16K_mp3_bokmaal_test_targets_no_lang_model.txt ADDED
The diff for this file is too large to render. See raw diff