infinitejoy commited on
Commit
50552ee
1 Parent(s): 492c67f

Upload eval.py

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