ayameRushia commited on
Commit
1e97e95
1 Parent(s): 9ae1879

Upload eval.py

Browse files
Files changed (1) hide show
  1. eval.py +158 -0
eval.py ADDED
@@ -0,0 +1,158 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from datasets import load_dataset, load_metric, Audio, Dataset
2
+ from transformers import pipeline, AutoFeatureExtractor, AutoTokenizer, AutoConfig, AutoModelForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM
3
+ import re
4
+ import torch
5
+ import argparse
6
+ from typing import Dict
7
+
8
+ def log_results(result: Dataset, args: Dict[str, str]):
9
+ """ DO NOT CHANGE. This function computes and logs the result metrics. """
10
+
11
+ log_outputs = args.log_outputs
12
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
13
+
14
+ # load metric
15
+ wer = load_metric("wer")
16
+ cer = load_metric("cer")
17
+
18
+ # compute metrics
19
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
20
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
21
+
22
+ # print & log results
23
+ result_str = (
24
+ f"WER: {wer_result}\n"
25
+ f"CER: {cer_result}"
26
+ )
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 = [",", "?", "¿", ".", "!", "¡", ";", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
53
+ "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
54
+ "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
55
+ "、", "﹂", "﹁", "‧", "~", "﹏", ",", "{", "}", "(", ")", "[", "]", "【", "】", "‥", "〽",
56
+ "『", "』", "〝", "〟", "⟨", "⟩", "〜", ":", "!", "?", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
57
+ chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
58
+
59
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
60
+
61
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
62
+ # note that order is important here!
63
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
64
+
65
+ for t in token_sequences_to_ignore:
66
+ text = " ".join(text.split(t))
67
+
68
+ return text
69
+
70
+
71
+ def main(args):
72
+ # load dataset
73
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
74
+
75
+ # for testing: only process the first two examples as a test
76
+ # dataset = dataset.select(range(10))
77
+
78
+ # load processor
79
+ if args.greedy:
80
+ processor = Wav2Vec2Processor.from_pretrained(args.model_id)
81
+ decoder = None
82
+ else:
83
+ processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
84
+ decoder = processor.decoder
85
+
86
+ feature_extractor = processor.feature_extractor
87
+ tokenizer = processor.tokenizer
88
+
89
+ # resample audio
90
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=feature_extractor.sampling_rate))
91
+
92
+ # load eval pipeline
93
+ if args.device is None:
94
+ args.device = 0 if torch.cuda.is_available() else -1
95
+
96
+ config = AutoConfig.from_pretrained(args.model_id)
97
+ model = AutoModelForCTC.from_pretrained(args.model_id)
98
+
99
+ #asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
100
+ asr = pipeline("automatic-speech-recognition", config=config, model=model, tokenizer=tokenizer,
101
+ feature_extractor=feature_extractor, decoder=decoder, device=args.device)
102
+
103
+
104
+ # map function to decode audio
105
+ def map_to_pred(batch, args=args, asr=asr):
106
+ prediction = asr(batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s)
107
+
108
+ batch["prediction"] = prediction["text"]
109
+ batch["target"] = normalize_text(batch["sentence"])
110
+ return batch
111
+
112
+ # run inference on all examples
113
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
114
+
115
+ # filtering out empty targets
116
+ result = result.filter(lambda example: example["target"] != "")
117
+
118
+ # compute and log_results
119
+ # do not change function below
120
+ log_results(result, args)
121
+
122
+
123
+ if __name__ == "__main__":
124
+ parser = argparse.ArgumentParser()
125
+
126
+ parser.add_argument(
127
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
128
+ )
129
+ parser.add_argument(
130
+ "--dataset", type=str, required=True, help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets"
131
+ )
132
+ parser.add_argument(
133
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
134
+ )
135
+ parser.add_argument(
136
+ "--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`"
137
+ )
138
+ parser.add_argument(
139
+ "--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."
140
+ )
141
+ parser.add_argument(
142
+ "--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."
143
+ )
144
+ parser.add_argument(
145
+ "--log_outputs", action='store_true', help="If defined, write outputs to log file for analysis."
146
+ )
147
+ parser.add_argument(
148
+ "--greedy", action='store_true', help="If defined, the LM will be ignored during inference."
149
+ )
150
+ parser.add_argument(
151
+ "--device",
152
+ type=int,
153
+ default=None,
154
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
155
+ )
156
+ args = parser.parse_args()
157
+
158
+ main(args)