eval.py
5.0 KB · 140 lines · python Raw
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
56 # In addition, we can normalize the target text, e.g. removing new lines characters etc...
57 # note that order is important here!
58 token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
59
60 for t in token_sequences_to_ignore:
61 text = " ".join(text.split(t))
62
63 #Replace some characters that are not in the Romanian alphabet but are found in the dataset
64 text = re.sub('[á]', 'a', text)
65 text = re.sub('[é]', 'e', text)
66 text = re.sub('[í]', 'i', text)
67 text = re.sub('[ò]', 'o', text)
68 text = re.sub('[ü]', 'u', text)
69 text = re.sub('[ć]', 'c', text)
70 text = re.sub('[đ]', 'd', text)
71 text = re.sub('[č]', 'c', text)
72 text = re.sub('[š]', 's', text)
73 text = re.sub('[ş]', 'ș', text) #Replace the s-cedilla that is sometimes (wrongly) used instead of the s-comma
74 text = re.sub('[ţ]', 'ț', text) #Replace the t-cedilla that is sometimes (wrongly) used instead of the t-comma
75 return text
76
77
78 def main(args):
79 # load dataset
80 dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
81
82 # for testing: only process the first two examples as a test
83 # dataset = dataset.select(range(10))
84
85 # load processor
86 feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
87 sampling_rate = feature_extractor.sampling_rate
88
89 # resample audio
90 dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
91
92 # load eval pipeline
93 asr = pipeline("automatic-speech-recognition", model=args.model_id)
94
95 # map function to decode audio
96 def map_to_pred(batch):
97 prediction = asr(
98 batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
99 )
100
101 batch["prediction"] = prediction["text"]
102 batch["target"] = normalize_text(batch["sentence"])
103 return batch
104
105 # run inference on all examples
106 result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
107
108 # compute and log_results
109 # do not change function below
110 log_results(result, args)
111
112
113 if __name__ == "__main__":
114 parser = argparse.ArgumentParser()
115
116 parser.add_argument(
117 "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
118 )
119 parser.add_argument(
120 "--dataset",
121 type=str,
122 required=True,
123 help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
124 )
125 parser.add_argument(
126 "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
127 )
128 parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
129 parser.add_argument(
130 "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
131 )
132 parser.add_argument(
133 "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
134 )
135 parser.add_argument(
136 "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
137 )
138 args = parser.parse_args()
139
140 main(args)