eval.py
4.7 KB · 141 lines · python Raw
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 = '[,?.!-;:""%\'"�\'‘’_，！łńō–—\\\\\\“”\\[\\]]'
54
55 text = re.sub(chars_to_ignore_regex, "", text.lower())
56 text = re.sub(r'[‘’´`]', r"'", text)
57 text = re.sub(r'è', r"é", text)
58 text = re.sub(r"(-|' | '| +)", " ", text)
59
60 # In addition, we can normalize the target text, e.g. removing new lines characters etc...
61 # note that order is important here!
62 token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
63
64 for t in token_sequences_to_ignore:
65 text = " ".join(text.split(t))
66
67 return text
68
69
70 def main(args):
71 # load dataset
72 dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
73
74 # for testing: only process the first two examples as a test
75 # dataset = dataset.select(range(10))
76
77 # load processor
78 feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
79 sampling_rate = feature_extractor.sampling_rate
80
81 # resample audio
82 dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
83
84 # load eval pipeline
85 if args.device is None:
86 args.device = 0 if torch.cuda.is_available() else -1
87 asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
88
89 # map function to decode audio
90 def map_to_pred(batch):
91 prediction = asr(
92 batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
93 )
94
95 batch["prediction"] = prediction["text"]
96 batch["target"] = normalize_text(batch["sentence"])
97 return batch
98
99 # run inference on all examples
100 result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
101
102 # compute and log_results
103 # do not change function below
104 log_results(result, args)
105
106
107 if __name__ == "__main__":
108 parser = argparse.ArgumentParser()
109
110 parser.add_argument(
111 "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
112 )
113 parser.add_argument(
114 "--dataset",
115 type=str,
116 required=True,
117 help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
118 )
119 parser.add_argument(
120 "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
121 )
122 parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
123 parser.add_argument(
124 "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
125 )
126 parser.add_argument(
127 "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
128 )
129 parser.add_argument(
130 "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
131 )
132 parser.add_argument(
133 "--device",
134 type=int,
135 default=None,
136 help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
137 )
138 args = parser.parse_args()
139
140 main(args)
141