deepdml commited on
Commit
15e94f2
1 Parent(s): 298d170

Create eval.py

Browse files
Files changed (1) hide show
  1. eval.py +144 -0
eval.py ADDED
@@ -0,0 +1,144 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 = [",", "?", "¿", ".", "!", "¡", ";", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
54
+ "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
55
+ "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
56
+ "、", "﹂", "﹁", "‧", "~", "﹏", ",", "{", "}", "(", ")", "[", "]", "【", "】", "‥", "〽",
57
+ "『", "』", "〝", "〟", "⟨", "⟩", "〜", ":", "!", "?", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"] # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
58
+
59
+ chars_to_remove_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
60
+
61
+ text = re.sub(chars_to_remove_regex, "", text.lower())
62
+ text = re.sub("[-]", " ", text)
63
+
64
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
65
+ # note that order is important here!
66
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
67
+
68
+ for t in token_sequences_to_ignore:
69
+ text = " ".join(text.split(t))
70
+
71
+ return text
72
+
73
+
74
+ def main(args):
75
+ # load dataset
76
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
77
+
78
+ # for testing: only process the first two examples as a test
79
+ # dataset = dataset.select(range(10))
80
+
81
+ # load processor
82
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
83
+ sampling_rate = feature_extractor.sampling_rate
84
+
85
+ # resample audio
86
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
87
+
88
+ # load eval pipeline
89
+ if args.device is None:
90
+ args.device = 0 if torch.cuda.is_available() else -1
91
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
92
+
93
+ # map function to decode audio
94
+ def map_to_pred(batch):
95
+ prediction = asr(
96
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
97
+ )
98
+
99
+ batch["prediction"] = prediction["text"]
100
+ batch["target"] = normalize_text(batch["sentence"])
101
+ return batch
102
+
103
+ # run inference on all examples
104
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
105
+
106
+ # compute and log_results
107
+ # do not change function below
108
+ log_results(result, args)
109
+
110
+
111
+ if __name__ == "__main__":
112
+ parser = argparse.ArgumentParser()
113
+
114
+ parser.add_argument(
115
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
116
+ )
117
+ parser.add_argument(
118
+ "--dataset",
119
+ type=str,
120
+ required=True,
121
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
122
+ )
123
+ parser.add_argument(
124
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
125
+ )
126
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
127
+ parser.add_argument(
128
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
129
+ )
130
+ parser.add_argument(
131
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
132
+ )
133
+ parser.add_argument(
134
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
135
+ )
136
+ parser.add_argument(
137
+ "--device",
138
+ type=int,
139
+ default=None,
140
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
141
+ )
142
+ args = parser.parse_args()
143
+
144
+ main(args)