File size: 10,727 Bytes
3af58b3
 
 
 
 
 
 
 
d3d8955
3af58b3
 
 
 
84a3664
0d4a063
3af58b3
 
 
 
 
 
 
0d4a063
 
 
 
 
3af58b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dadf62a
eae9a48
3af58b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb05670
3af58b3
 
bb05670
 
 
0d4a063
 
 
 
 
 
 
 
 
bb05670
 
0d4a063
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3af58b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb05670
3af58b3
 
 
 
 
 
 
 
0d4a063
3af58b3
a64382f
 
 
a1e8de0
3af58b3
 
 
 
 
 
 
 
a64382f
3af58b3
 
 
 
 
1567496
 
 
6e3e8e8
1567496
 
3af58b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
df9222e
 
3af58b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1567496
d3d8955
1567496
3af58b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
df9222e
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
#!/usr/bin/env python3
import argparse
import re
from typing import Dict

import torch
from datasets import Audio, Dataset, load_dataset, load_metric
from num2words import num2words as n2w
from slugify import slugify

from transformers import AutoFeatureExtractor, AutoModelForCTC, pipeline, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM, Wav2Vec2FeatureExtractor
# from pyctcdecode import BeamSearchDecoderCTC

from cardinal_numbers import convert_nums


def log_results(result: Dataset, args: Dict[str, str]):
    """DO NOT CHANGE. This function computes and logs the result metrics."""

    log_outputs = args.log_outputs
    lm = "withLM" if args.use_lm else "noLM"
    model_id = args.model_id.replace("/", "_").replace(".", "")
    if args.filter:
        extra_args = [args.config, slugify(args.filter), args.split, lm]
    else:
        extra_args = [args.config, args.split, lm]
    dataset_id = "_".join([model_id] + args.dataset.split("/") + extra_args)

    # load metric
    wer = load_metric("wer")
    cer = load_metric("cer")

    # compute metrics
    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])

    # print & log results
    result_str = f"{dataset_id}\nWER: {wer_result}\nCER: {cer_result}"
    print(result_str)

    with open(f"{dataset_id}_eval_results.txt", "w") as f:
        f.write(result_str)
    with open(f"{dataset_id}_eval_results.tsv", "w") as f:
        f.write("\t".join([args.model_id, args.dataset, args.config, args.filter, args.split, str(lm), str(wer_result), str(cer_result)]))

    # log all results in text file. Possibly interesting for analysis
    if log_outputs is not None:
        pred_file = f"log_{dataset_id}_predictions.txt"
        target_file = f"log_{dataset_id}_targets.txt"

        with open(pred_file, "w") as p, open(target_file, "w") as t:
            # mapping function to write output
            def write_to_file(batch, i):
                p.write(f"{i}" + "\n")
                p.write(batch["prediction"] + "\n")
                t.write(f"{i}" + "\n")
                t.write(batch["target"] + "\n")

            result.map(write_to_file, with_indices=True)


def normalize_text(original_text: str, dataset: str) -> str:
    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""

    text = original_text.lower()
    if dataset.lower().endswith("fleurs"): 
        replacements = (
            (r"\be\.kr", "etter kristus fødsel"),
            (r"\bf\.kr", "før kristi fødsel"),
            (r"\bca[.]?\b", "circa"),
            (r"(\d)\s*km/t", r"\1 kilometer i timen"),
            (r"(\d)\s*km", r"\1 kilometer"),
            (r"(\d)\s*cm", r"\1 centimeter"),
            (r"(\d)\s*mm", r"\1 millimeter"),
            (r"kl\.", "klokka"),
            (r"f\.eks", "for eksempel"),
        )
        for abrev, expasion in replacements:
            text = re.sub(abrev, expasion, text)
        text = re.sub(r'(\d+)[-–](\d+)', r'\1 til \2', text)  # 1-89, 70-90
        text = re.sub(r'(\d{2}):00', r'\1', text)  # 21:00
        text = re.sub(r"(\d{2}):0(\d{1})", r"\1 null \2", text)  # 17:03
        text = re.sub(r"(\d{1,2}):(\d{1,2})", r"\1 \2", text)  # 17:23 (time), 4:3 (aspect ratios)
        text = re.sub(r"(1[1-9])00", r"\1 hundre", text)  # 1800, 1900
        text = re.sub(r"(1[1-9])0([1-9])", r"\1 null \2 ", text)  # 1901, 1909
        text = re.sub(r"(1[1-9])([1-9]\d)", r"\1 \2 ", text)  # 1911, 1987
        text = re.sub(r"(20)0([1-9])", r"\1 null \2 ", text)  # 2009
        text = re.sub(r"(20)(\d{2})", r"\1 \2 ", text)  # 2009
        text = re.sub(r"(\d{1,3})[.](\d{1,2})", r"\1 dot \2 ", text)  # 802.11n, 2.5ghz (in English)
        text = re.sub(r"(\d{1,2})[ .](\d{3})", r"\1\2", text)  # 10 000, 32.000
        text = re.sub(r'(\w+)-(\w+)', r'\1 \2', text)  # n-standard
        # text = re.compile(r"-?0?[1-9][\d.]*").sub(lambda x: n2w(x.group(0), lang="no"), text.replace(".", ""))
        text = re.compile(r"-?0?[1-9][\d.]*").sub(lambda x: convert_nums(int(x.group(0)), nn=True), text.replace(".", ""))


    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
    text = re.sub(chars_to_ignore_regex, "", text) + " "

    if dataset.lower().endswith("nst"):
        text = text.lower()
        text = text.replace("(...vær stille under dette opptaket...)", "")
        text = re.sub('[áàâ]', 'a', text)
        text = re.sub('[ä]', 'æ', text)
        text = re.sub('[éèëê]', 'e', text)
        text = re.sub('[íìïî]', 'i', text)
        text = re.sub('[óòöô]', 'o', text)
        text = re.sub('[ö]', 'ø', text)
        text = re.sub('[ç]', 'c', text)
        text = re.sub('[úùüû]', 'u', text)
        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
        text = re.sub('\s+', ' ', text)
    elif dataset.lower().endswith("npsc"):
        text = re.sub('[áàâ]', 'a', text)
        text = re.sub('[ä]', 'æ', text)
        text = re.sub('[éèëê]', 'e', text)
        text = re.sub('[íìïî]', 'i', text)
        text = re.sub('[óòöô]', 'o', text)
        text = re.sub('[ö]', 'ø', text)
        text = re.sub('[ç]', 'c', text)
        text = re.sub('[úùüû]', 'u', text)
        text = re.sub('\s+', ' ', text)
    elif dataset.lower().endswith("fleurs"): 
        text = re.sub('[áàâ]', 'a', text)
        text = re.sub('[ä]', 'æ', text)
        text = re.sub('[éèëê]', 'e', text)
        text = re.sub('[íìïî]', 'i', text)
        text = re.sub('[óòöô]', 'o', text)
        text = re.sub('[ö]', 'ø', text)
        text = re.sub('[ç]', 'c', text)
        text = re.sub('[úùüû]', 'u', text)
        text = re.sub('[«»]', '', text)
        text = re.sub('\s+', ' ', text)
    text = re.sub('<ee>', 'eee', text)
    text = re.sub('<qq>', 'qqq', text)
    text = re.sub('<mm>', 'mmm', text)
    text = re.sub('<inaudible>', 'xxx', text)

    # # In addition, we can normalize the target text, e.g. removing new lines characters etc...
    # # note that order is important here!
    # token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]

    # for t in token_sequences_to_ignore:
    #     text = " ".join(text.split(t))

    return text


def main(args):
    # load dataset
    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
    if args.filter:
        attribute, value = list(map(str.strip, args.filter.split(":")))
        dataset = dataset.filter(
            lambda x: x[attribute] == value,
            desc=f"Filtering on {args.filter}",
        )
    # for testing: only process the first two examples as a test
    # dataset = dataset.select(range(10))

    # load processor
    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
    sampling_rate = feature_extractor.sampling_rate

    # resample audio
    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))

    # load eval pipeline
    if args.device is None:
        args.device = 0 if torch.cuda.is_available() else -1
    # asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)

    model_instance = AutoModelForCTC.from_pretrained(args.model_id)
    if args.use_lm:
        processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
        decoder = processor.decoder
    else:
        processor = Wav2Vec2Processor.from_pretrained(args.model_id)
        decoder = None
    asr = pipeline(
        "automatic-speech-recognition",
        model=model_instance,
        tokenizer=processor.tokenizer, 
        feature_extractor=processor.feature_extractor,
        decoder=decoder,
        device=args.device
    )

    # feature_extractor_dict, _ = Wav2Vec2FeatureExtractor.get_feature_extractor_dict(args.model_id)
    # feature_extractor_dict["processor_class"] = "Wav2Vec2Processor" if not args.use_lm else "Wav2Vec2ProcessorWithLM"
    # feature_extractor = Wav2Vec2FeatureExtractor.from_dict(feature_extractor_dict)

    # asr = pipeline("automatic-speech-recognition", model=args.model_id, feature_extractor=feature_extractor, device=args.device, decoder=BeamSearchDecoderCTC.load_from_dir("./"))

    # map function to decode audio
    def map_to_pred(batch):
        prediction = asr(
            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
        )

        batch["prediction"] = prediction["text"]
        batch["target"] = normalize_text(batch[args.text_column], args.dataset)
        return batch

    # run inference on all examples
    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)

    # compute and log_results
    # do not change function below
    log_results(result, args)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.add_argument(
        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
    )
    parser.add_argument(
        "--dataset",
        type=str,
        required=True,
        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
    )
    parser.add_argument(
        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
    )
    parser.add_argument(
        "--filter", type=str, default="", help="Simple filter on attributes. *E.g.* `region_of_youth:Troms` would pnly keep those samplesfor which the condition is met"
    )
    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
    parser.add_argument(
        "--text_column", type=str, default="text", help="Column name containing the transcription."
    )
    parser.add_argument(
        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
    )
    parser.add_argument(
        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
    )
    parser.add_argument(
        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
    )
    parser.add_argument(
        "--device",
        type=int,
        default=None,
        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
    )
    parser.add_argument(
        "--use_lm", action="store_true", help="If defined, use included language model as the decoder."
    )
    args = parser.parse_args()

    main(args)