File size: 6,534 Bytes
841ba8a
 
25b96d6
841ba8a
e42ea01
 
841ba8a
 
25b96d6
841ba8a
25b96d6
841ba8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e42ea01
841ba8a
e42ea01
 
 
 
 
 
 
 
 
841ba8a
 
 
 
 
 
 
 
 
 
 
25b96d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0cc9598
25b96d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
841ba8a
 
 
 
 
e42ea01
841ba8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25b96d6
841ba8a
 
 
 
 
 
 
25b96d6
 
 
 
 
 
 
 
 
 
 
 
 
0cc9598
841ba8a
0cc9598
841ba8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25b96d6
 
 
841ba8a
 
 
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
#!/usr/bin/env python3
import argparse
import functools
import re
import string
import unidecode
from typing import Dict

from datasets import Audio, Dataset, DatasetDict, load_dataset, load_metric

from transformers import AutoFeatureExtractor, AutoTokenizer, pipeline


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
    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])

    # 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"WER: {wer_result}\n" f"CER: {cer_result}"
    print(result_str)

    with open(f"{dataset_id}_eval_results.txt", "w") as f:
        f.write(result_str)

    # 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(text: str) -> str:
    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""

    chars_to_ignore_regex = f'[{re.escape(string.punctuation)}]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training

    text = re.sub(
        chars_to_ignore_regex, 
        "", 
        re.sub("['`´]", "’",   # elsewhere probably meant as glottal stop
               re.sub("([og])['`´]", "\g<1>‘",  # after o/g indicate modified char
                      unidecode.unidecode(text).lower()
                     )
              )
    ) + " "

    # 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 create_vocabulary_from_data(
    datasets: DatasetDict,
    word_delimiter_token = None,
    unk_token = None,
    pad_token = None,
):
    # Given training and test labels create vocabulary
    def extract_all_chars(batch):
        all_text = " ".join(batch["target"])
        vocab = list(set(all_text))
        return {"vocab": [vocab], "all_text": [all_text]}

    vocabs = datasets.map(
        extract_all_chars,
        batched=True,
        batch_size=-1,
        keep_in_memory=True,
        remove_columns=datasets["test"].column_names,
    )


    vocab_dict = {v: k for k, v in enumerate(sorted(vocabs["test"]["vocab"][0]))}

    # replace white space with delimiter token
    if word_delimiter_token is not None:
        vocab_dict[word_delimiter_token] = vocab_dict[" "]
        del vocab_dict[" "]

    # add unk and pad token
    if unk_token is not None:
        vocab_dict[unk_token] = len(vocab_dict)

    if pad_token is not None:
        vocab_dict[pad_token] = len(vocab_dict)

    return vocab_dict


def main(args):
    # load dataset
    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)

    # 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
    asr = pipeline("automatic-speech-recognition", model=args.model_id)

    # 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["sentence"])
        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 args.check_vocab:
        tokenizer = AutoTokenizer.from_pretrained(args.model_id)
        unk_token = "[UNK]"
        pad_token = "[PAD]"
        word_delimiter_token = "|"
        raw_datasets = DatasetDict({"test": result})
        vocab_dict = create_vocabulary_from_data(
                    raw_datasets,
                    word_delimiter_token=word_delimiter_token,
                    unk_token=unk_token,
                    pad_token=pad_token,
                )
        print(vocab_dict)
        print("OOV chars:", set(vocab_dict) - set(tokenizer.get_vocab()))    

        
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("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
    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(
        "--check_vocab", action="store_true", help="Verify that normalized target text is within character set"
    )
    args = parser.parse_args()

    main(args)