cahya commited on
Commit
d2fdc8c
1 Parent(s): 856a788

add eval_kenlm.py

Browse files
Files changed (1) hide show
  1. eval_kenlm.py +165 -0
eval_kenlm.py ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, Wav2Vec2ForCTC, Wav2Vec2Processor, set_seed
10
+ from pyctcdecode import build_ctcdecoder
11
+ from multiprocessing import Pool
12
+
13
+ class KenLM:
14
+ def __init__(self, tokenizer, model_name, unigrams=None, num_workers=8, beam_width=128):
15
+ self.num_workers = num_workers
16
+ self.beam_width = beam_width
17
+ vocab_dict = tokenizer.get_vocab()
18
+ self.vocabulary = [x[0] for x in sorted(vocab_dict.items(), key=lambda x: x[1], reverse=False)]
19
+ self.vocabulary = self.vocabulary[:-1]
20
+ self.decoder = build_ctcdecoder(self.vocabulary, model_name, unigrams=unigrams)
21
+
22
+ @staticmethod
23
+ def lm_postprocess(text):
24
+ return ' '.join([x if len(x) > 1 else "" for x in text.split()]).strip()
25
+
26
+ def decode(self, logits):
27
+ probs = logits.cpu().numpy()
28
+ # probs = logits.numpy()
29
+ with Pool(self.num_workers) as pool:
30
+ text = self.decoder.decode_batch(pool, probs)
31
+ text = [KenLM.lm_postprocess(x) for x in text]
32
+ return text
33
+
34
+ def log_results(result: Dataset, args: Dict[str, str]):
35
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
36
+
37
+ log_outputs = args.log_outputs
38
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
39
+
40
+ # load metric
41
+ wer = load_metric("wer")
42
+ cer = load_metric("cer")
43
+
44
+ # compute metrics
45
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
46
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
47
+
48
+ # print & log results
49
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
50
+ print(result_str)
51
+
52
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
53
+ f.write(result_str)
54
+
55
+ # log all results in text file. Possibly interesting for analysis
56
+ if log_outputs is not None:
57
+ pred_file = f"log_{dataset_id}_predictions.txt"
58
+ target_file = f"log_{dataset_id}_targets.txt"
59
+
60
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
61
+
62
+ # mapping function to write output
63
+ def write_to_file(batch, i):
64
+ p.write(f"{i}" + "\n")
65
+ p.write(batch["prediction"] + "\n")
66
+ t.write(f"{i}" + "\n")
67
+ t.write(batch["target"] + "\n")
68
+
69
+ result.map(write_to_file, with_indices=True)
70
+
71
+
72
+ def normalize_text(text: str) -> str:
73
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
74
+ chars_to_ignore_regex = '[,?.!-;:""%\'"\'\'`…’»«‘“”�éû]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
75
+
76
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
77
+
78
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
79
+ # note that order is important here!
80
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
81
+
82
+ for t in token_sequences_to_ignore:
83
+ text = " ".join(text.split(t))
84
+
85
+ return text
86
+
87
+
88
+ def main(args):
89
+ # load dataset
90
+ dataset = load_dataset(args.dataset, args.config, data_dir=args.data_dir, split=args.split, use_auth_token=True)
91
+
92
+ # for testing: only process the first two examples as a test
93
+ # dataset = dataset.select(range(10))
94
+
95
+ # load processor
96
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
97
+ sampling_rate = feature_extractor.sampling_rate
98
+
99
+ # resample audio
100
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
101
+
102
+ # load eval pipeline
103
+ if args.device is None:
104
+ args.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
105
+
106
+ set_seed(42) # set the random seed to have reproducible result.
107
+ processor = Wav2Vec2Processor.from_pretrained(args.model_id)
108
+ model = Wav2Vec2ForCTC.from_pretrained(args.model_id)
109
+ kenlm = KenLM(processor.tokenizer, "language_model/5gram.bin", unigrams="language_model/unigrams.txt")
110
+
111
+ # map function to decode audio
112
+ def map_to_pred(batch):
113
+ inputs = processor(batch["audio"]["array"], sampling_rate=16_000, return_tensors="pt", padding=True)
114
+ with torch.no_grad():
115
+ logits = model(inputs.input_values.to(args.device), attention_mask=inputs.attention_mask.to(args.device)).logits
116
+ prediction = kenlm.decode(logits)
117
+
118
+ batch["prediction"] = prediction
119
+ batch["target"] = normalize_text(batch["sentence"])
120
+ return batch
121
+
122
+ # run inference on all examples
123
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
124
+
125
+ # compute and log_results
126
+ # do not change function below
127
+ log_results(result, args)
128
+
129
+
130
+ if __name__ == "__main__":
131
+ parser = argparse.ArgumentParser()
132
+
133
+ parser.add_argument(
134
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
135
+ )
136
+ parser.add_argument(
137
+ "--dataset",
138
+ type=str,
139
+ required=True,
140
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
141
+ )
142
+ parser.add_argument(
143
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
144
+ )
145
+ parser.add_argument("--data_dir", type=str, required=False, default=None,
146
+ help="The directory contains the dataset")
147
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
148
+ parser.add_argument(
149
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
150
+ )
151
+ parser.add_argument(
152
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
153
+ )
154
+ parser.add_argument(
155
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
156
+ )
157
+ parser.add_argument(
158
+ "--device",
159
+ type=int,
160
+ default=None,
161
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
162
+ )
163
+ args = parser.parse_args()
164
+
165
+ main(args)