Baybars commited on
Commit
56c3efb
1 Parent(s): 15ec2be

eval and language model added

Browse files
README.md CHANGED
@@ -38,6 +38,9 @@ More information needed
38
 
39
  More information needed
40
 
 
 
 
41
  ## Training procedure
42
 
43
  ### Training hyperparameters
 
38
 
39
  More information needed
40
 
41
+ ## Language Model
42
+ N-gram language model is trained by [mpoyraz](https://huggingface.co/mpoyraz/wav2vec2-xls-r-300m-cv7-turkish) on a Turkish Wikipedia articles using KenLM and [ngram-lm-wiki](https://github.com/mpoyraz/ngram-lm-wiki) repo was used to generate arpa LM and convert it into binary format.
43
+
44
  ## Training procedure
45
 
46
  ### Training hyperparameters
eval.py ADDED
@@ -0,0 +1,151 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ from unicode_tr import unicode_tr
12
+
13
+ def log_results(result: Dataset, args: Dict[str, str]):
14
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
15
+
16
+ log_outputs = args.log_outputs
17
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
18
+
19
+ # load metric
20
+ wer = load_metric("wer")
21
+ cer = load_metric("cer")
22
+
23
+ # compute metrics
24
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
25
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
26
+
27
+ # print & log results
28
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
29
+ print(result_str)
30
+
31
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
32
+ f.write(result_str)
33
+
34
+ # log all results in text file. Possibly interesting for analysis
35
+ if log_outputs is not None:
36
+ pred_file = f"log_{dataset_id}_predictions.txt"
37
+ target_file = f"log_{dataset_id}_targets.txt"
38
+
39
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
40
+
41
+ # mapping function to write output
42
+ def write_to_file(batch, i):
43
+ p.write(f"{i}" + "\n")
44
+ p.write(batch["prediction"] + "\n")
45
+ t.write(f"{i}" + "\n")
46
+ t.write(batch["target"] + "\n")
47
+
48
+ result.map(write_to_file, with_indices=True)
49
+
50
+
51
+ def normalize_text(text: str) -> str:
52
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
53
+
54
+ chars_to_remove_regex = '[,?.!\-\;\:"“%”�—…–()]'
55
+ apostrophes = "[’‘`´ʹʻʼʽʿˈ]"
56
+
57
+ # Lower the text using 'unicode_tr'
58
+ # Regular lower() does not work well for Turkish Language
59
+ text_norm = unicode_tr(text).lower()
60
+ # Unify apostrophes
61
+ text_norm = re.sub(apostrophes, "'", text_norm)
62
+ # Remove pre-defined chars
63
+ text_norm = re.sub(chars_to_remove_regex, "", text_norm)
64
+ # Remove single quotes
65
+ text_norm = text_norm.replace(" '", " ")
66
+ text_norm = text_norm.replace("' ", " ")
67
+ # Handle hatted characters
68
+ text_norm = re.sub('[â]', 'a', text_norm)
69
+ text_norm = re.sub('[î]', 'i', text_norm)
70
+ text_norm = re.sub('[ô]', 'o', text_norm)
71
+ text_norm = re.sub('[û]', 'u', text_norm)
72
+ # Handle alternate characters
73
+ text_norm = re.sub('[é]', 'e', text_norm)
74
+ text_norm = re.sub('[ë]', 'e', text_norm)
75
+ # Remove multiple spaces
76
+ text_norm = re.sub(r"\s+", " ", text_norm)
77
+
78
+ return text_norm
79
+
80
+
81
+ def main(args):
82
+ # load dataset
83
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
84
+
85
+ # for testing: only process the first two examples as a test
86
+ # dataset = dataset.select(range(10))
87
+
88
+ # load processor
89
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
90
+ sampling_rate = feature_extractor.sampling_rate
91
+
92
+ # resample audio
93
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
94
+
95
+ # load eval pipeline
96
+ if args.device is None:
97
+ args.device = 0 if torch.cuda.is_available() else -1
98
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
99
+
100
+ # map function to decode audio
101
+ def map_to_pred(batch):
102
+ prediction = asr(
103
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
104
+ )
105
+
106
+ batch["prediction"] = prediction["text"]
107
+ batch["target"] = normalize_text(batch["sentence"])
108
+ return batch
109
+
110
+ # run inference on all examples
111
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
112
+
113
+ # compute and log_results
114
+ # do not change function below
115
+ log_results(result, args)
116
+
117
+
118
+ if __name__ == "__main__":
119
+ parser = argparse.ArgumentParser()
120
+
121
+ parser.add_argument(
122
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
123
+ )
124
+ parser.add_argument(
125
+ "--dataset",
126
+ type=str,
127
+ required=True,
128
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
129
+ )
130
+ parser.add_argument(
131
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
132
+ )
133
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
134
+ parser.add_argument(
135
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
136
+ )
137
+ parser.add_argument(
138
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
139
+ )
140
+ parser.add_argument(
141
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
142
+ )
143
+ parser.add_argument(
144
+ "--device",
145
+ type=int,
146
+ default=None,
147
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
148
+ )
149
+ args = parser.parse_args()
150
+
151
+ main(args)
language_model/attrs.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}
language_model/lm.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ca3b400bd46dd68a7999b10c4ff87aa79211e48f22e56c68a45c18fcf4387971
3
+ size 496120504
language_model/unigrams.txt ADDED
The diff for this file is too large to render. See raw diff