|
|
|
from datasets import load_dataset, load_metric, Audio, Dataset |
|
from transformers import pipeline, AutoFeatureExtractor, AutoTokenizer, AutoConfig, AutoModelForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM |
|
import re |
|
import torch |
|
import argparse |
|
from typing import Dict |
|
|
|
|
|
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]) |
|
|
|
|
|
wer = load_metric("wer") |
|
cer = load_metric("cer") |
|
|
|
|
|
wer_result = wer.compute(references=result["target"], predictions=result["prediction"]) |
|
cer_result = cer.compute(references=result["target"], predictions=result["prediction"]) |
|
|
|
|
|
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) |
|
|
|
|
|
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: |
|
|
|
|
|
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, invalid_chars_regex: str, to_lower: bool) -> str: |
|
|
|
chars_to_ignore_regex = '[\µ\я\ひ\ⱎ\ⱅ\ḥ\ӌ\џ\ŵ\ʋ\λ\φ\χ\г\и\к\п\ц\ч\э\я\џ\ӌ\ቀ\ከ\ጀ\ḥ\牡\津\宇\厳\保\丹\三\む\も\ⱎ\ⱅ\⋅\⊨\↔\ℚ\э\п\к\и\г\|\£\§\·\½\º\ə\ơ\ǀ\ː\ʾ\ˢ\г\и\к\э\п\∨\„\,\?\.\!\—\―\–\;\:\"\‘\»\%\ł\_\€\×\ぬ\;\±\ß\Þ\«\Ø\°\…\”\“\`\ʿ\&\=\+\の\~\(\)\Σ\ı\ጠ\ℵ\馆\青\貴\西\美\甌\杜\术\星\文\扬\北\京\乃\ゔ\や\め\ま\へ\つ\た\う\い\☉\≥\®\/\∞\∆\∅\→\ℰ\ω\ψ\Μ\Θ\Κ\Π\Σ\Ω\α\γ\δ\ε\ζ\η\κ\ι\ν\μ\ρ\ς\σ\τ\υ\ℤ\ℝ\ℂ\ℕ\₽\∈\›\ο\‹\†\}\{\}\_\ደ\Δ\ወ\ي\و\ب\ة\د\ن\ن\ل\را\э\р\п\н\м\к\и\з\ψ\υ\θ\ṭ\ṯ\ḍ\*\^\∼\م\э\п\ǃ\$\Ꝑ]' |
|
chars_to_replace_a = '[\ɑ\ạ\ả\ầ\ậ\ắ\ẵ\а\ǎ\ā\ă\ą\á\ã\ä\å]' |
|
chars_to_replace_i = '[\ɨ\ị\ı\ī\ĩ\í\ì\і]' |
|
chars_to_replace_e = '[\ệ\ễ\ề\ě\ę\ė\ē\е\ế]' |
|
chars_to_replace_o = '[\ồ\ộ\ờ\ợ\ő\ö\ŏ\ō\ø\õ\ó\ò\ð\ǫ\ό\ớ\ổ\ố]' |
|
chars_to_replace_u = '[\ų\ʉ\ủ\ử\ù\ü\ư\ǔ\ů\ū\ũ\ú\ứ\ụ\ű\ŭ]' |
|
chars_to_replace_c = '[\ς\ć\ċ\č\ҫ]' |
|
chars_to_replace_y = '[\ÿ\ỳ\ÿ\ý]' |
|
chars_to_replace_n = '[\ṇ\ṅ\ǹ\ħ\ñ\ň\ņ\ń]' |
|
chars_to_replace_t = '[\ṭ\ț\ť\ţ]' |
|
chars_to_replace_s = '[\ṣ\ș\š\ş\ś]' |
|
chars_to_replace_q = '[\զ\գ\գ\զ]' |
|
chars_to_replace_j = '[\ј]' |
|
chars_to_replace_z = '[\ž\ż\ź\ẓ]' |
|
chars_to_replace_r = '[\ř]' |
|
chars_to_replace_l = '[\ł\ļ\ĺ]' |
|
chars_to_replace_k = '[\ķ]' |
|
chars_to_replace_g = '[\ġ\ğ]' |
|
chars_to_replace_d = '[\đ\ď]' |
|
chars_to_replace_b = '[\þ]' |
|
chars_to_replace_p = '[\р]' |
|
chars_to_replace_apostrophe = '[\´\′\ʼ\’\'\'\ʽ\ʻ\ʾ]' |
|
chars_to_replace_tirets = '[\─\−\‐]' |
|
|
|
|
|
if to_lower: |
|
text = re.sub(chars_to_ignore_regex, " ", text).lower() |
|
text = re.sub(chars_to_replace_a, "a", text) |
|
text = re.sub(chars_to_replace_i, "i", text) |
|
text = re.sub(chars_to_replace_e, "e", text) |
|
text = re.sub(chars_to_replace_o, "o", text) |
|
text = re.sub(chars_to_replace_u, "u", text) |
|
text = re.sub(chars_to_replace_c, "c", text) |
|
text = re.sub(chars_to_replace_y, "y", text) |
|
text = re.sub(chars_to_replace_n, "n", text) |
|
text = re.sub(chars_to_replace_t, "t", text) |
|
text = re.sub(chars_to_replace_s, "s", text) |
|
text = re.sub(chars_to_replace_q, "q", text) |
|
text = re.sub(chars_to_replace_j, "j", text) |
|
text = re.sub(chars_to_replace_z, "z", text) |
|
text = re.sub(chars_to_replace_r, "r", text) |
|
text = re.sub(chars_to_replace_l, "l", text) |
|
text = re.sub(chars_to_replace_k, "k", text) |
|
text = re.sub(chars_to_replace_g, "g", text) |
|
text = re.sub(chars_to_replace_d, "d", text) |
|
text = re.sub(chars_to_replace_b, "b", text) |
|
text = re.sub(chars_to_replace_q, "q", text) |
|
text = re.sub(chars_to_replace_p, "p", text) |
|
text = re.sub(chars_to_replace_apostrophe, "'", text) |
|
text = re.sub(chars_to_replace_tirets, "-", text) |
|
text = re.sub("β", "beta", text) |
|
text = re.sub("æ", "ae", text) |
|
text = re.sub("œ", "oe", text) |
|
text = re.sub("&", "et", text) |
|
text = re.sub("π", "pi", text) |
|
text = re.sub("ľ", "l'", text) |
|
text = re.sub(r"^\s+|\s+$", "", text) |
|
text = re.sub(" +", " ", text) |
|
text = re.sub("\n", " ", text) |
|
return text |
|
|
|
|
|
def main(args): |
|
|
|
dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True) |
|
|
|
|
|
|
|
|
|
|
|
if args.greedy: |
|
processor = Wav2Vec2Processor.from_pretrained(args.model_id) |
|
decoder = None |
|
else: |
|
processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id) |
|
decoder = processor.decoder |
|
|
|
|
|
feature_extractor = processor.feature_extractor |
|
tokenizer = processor.tokenizer |
|
|
|
|
|
dataset = dataset.cast_column("audio", Audio(sampling_rate=feature_extractor.sampling_rate)) |
|
|
|
|
|
if args.device is None: |
|
args.device = 0 if torch.cuda.is_available() else -1 |
|
|
|
config = AutoConfig.from_pretrained(args.model_id) |
|
model = AutoModelForCTC.from_pretrained(args.model_id) |
|
|
|
|
|
asr = pipeline("automatic-speech-recognition", config=config, model=model, tokenizer=tokenizer, |
|
feature_extractor=feature_extractor, decoder=decoder, device=args.device) |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(args.model_id) |
|
tokens = [x for x in tokenizer.convert_ids_to_tokens(range(0, tokenizer.vocab_size))] |
|
special_tokens = [ |
|
tokenizer.pad_token, tokenizer.word_delimiter_token, |
|
tokenizer.unk_token, tokenizer.bos_token, |
|
tokenizer.eos_token, |
|
] |
|
non_special_tokens = [x for x in tokens if x not in special_tokens] |
|
invalid_chars_regex = f"[^\s{re.escape(''.join(set(non_special_tokens)))}]" |
|
normalize_to_lower = False |
|
for token in non_special_tokens: |
|
if token.isalpha() and token.islower(): |
|
normalize_to_lower = True |
|
break |
|
|
|
|
|
def map_to_pred(batch, args=args, asr=asr, invalid_chars_regex=invalid_chars_regex, normalize_to_lower=normalize_to_lower): |
|
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"], invalid_chars_regex, normalize_to_lower) |
|
return batch |
|
|
|
|
|
result = dataset.map(map_to_pred, remove_columns=dataset.column_names) |
|
|
|
|
|
|
|
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( |
|
"--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 None. For long audio files a good value would be 5.0 seconds." |
|
) |
|
parser.add_argument( |
|
"--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." |
|
) |
|
parser.add_argument( |
|
"--log_outputs", action='store_true', help="If defined, write outputs to log file for analysis." |
|
) |
|
parser.add_argument( |
|
"--greedy", action='store_true', help="If defined, the LM will be ignored during inference." |
|
) |
|
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.", |
|
) |
|
args = parser.parse_args() |
|
|
|
main(args) |
