AndrewMcDowell commited on
Commit
561be41
1 Parent(s): 44deea1

eval results.

Browse files
eval.py ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ from pykakasi import kakasi
55
+
56
+ kakasi = kakasi()
57
+ kakasi.setMode('J', 'H') #Convert from kanji to hiragana
58
+ conv = kakasi.getConverter()
59
+ chars_to_ignore_regex = '[\,\?\!\-\;\:\"\“\%\‘\”\�\—\’\…\–\(\,\[\]\)\(\!]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
60
+
61
+ text = text.lower()
62
+ # normalize non-standard (stylized) unicode characters
63
+ text = unicodedata.normalize('NFKC', text)
64
+ # remove punctuation
65
+ text = conv.do(re.sub(chars_to_ignore_regex, "", text))
66
+
67
+ # Let's also make sure we split on all kinds of newlines, spaces, etc...
68
+ text = " ".join(text.split())
69
+
70
+ return text
71
+
72
+
73
+ def main(args):
74
+ # load dataset
75
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
76
+
77
+ # for testing: only process the first two examples as a test
78
+ # dataset = dataset.select(range(10))
79
+
80
+ # load processor
81
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
82
+ sampling_rate = feature_extractor.sampling_rate
83
+
84
+ # resample audio
85
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
86
+
87
+ # load eval pipeline
88
+ asr = pipeline("automatic-speech-recognition", model=args.model_id)
89
+
90
+ # map function to decode audio
91
+ def map_to_pred(batch):
92
+ prediction = asr(batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s)
93
+
94
+ batch["prediction"] = prediction["text"]
95
+ batch["target"] = normalize_text(batch["sentence"])
96
+ return batch
97
+
98
+ # run inference on all examples
99
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
100
+
101
+ # compute and log_results
102
+ # do not change function below
103
+ log_results(result, args)
104
+
105
+
106
+ if __name__ == "__main__":
107
+ parser = argparse.ArgumentParser()
108
+
109
+ parser.add_argument(
110
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
111
+ )
112
+ parser.add_argument(
113
+ "--dataset", type=str, required=True, help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets"
114
+ )
115
+ parser.add_argument(
116
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
117
+ )
118
+ parser.add_argument(
119
+ "--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`"
120
+ )
121
+ parser.add_argument(
122
+ "--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."
123
+ )
124
+ parser.add_argument(
125
+ "--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."
126
+ )
127
+ parser.add_argument(
128
+ "--log_outputs", action='store_true', help="If defined, write outputs to log file for analysis."
129
+ )
130
+ args = parser.parse_args()
131
+
132
+ main(args)
log_mozilla-foundation_common_voice_8_0_ja_test_predictions.txt ADDED
The diff for this file is too large to render. See raw diff
 
log_mozilla-foundation_common_voice_8_0_ja_test_targets.txt ADDED
The diff for this file is too large to render. See raw diff
 
mozilla-foundation_common_voice_8_0_ja_test_eval_results.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ WER: 0.9855427046263345
2
+ CER: 0.328342726455934