telugu_eval.py

#!/usr/bin/env python3
import torch    
import re
import argparse

from datasets import load_dataset, load_metric, Audio, Dataset
from transformers import pipeline, AutoFeatureExtractor, Wav2Vec2ProcessorWithLM, Wav2Vec2Processor
from transformers import Wav2Vec2ForCTC, AutoModelForCTC, AutoProcessor
from typing import Dict


def log_results(result: Dataset, args: Dict[str, str]):

    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}"
    )

    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:

            # 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 remove_special_characters(batch):
    chars_to_remove_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\&\/\d\_\\\]'
    batch["sentence"] = re.sub(chars_to_remove_regex, '', batch["sentence"]).lower()
    batch["sentence"] = re.sub('\u200c', '', batch["sentence"])
    batch["sentence"] = re.sub('[a-z]', '', batch["sentence"])
    return batch

def main(args):
    # load dataset
    dataset = load_dataset(args.dataset, args.config)
    train_testvalid = dataset[args.split].train_test_split(test_size=0.25)
    dataset_train = train_testvalid["train"]
    dataset_test = train_testvalid["test"]

    # load processor
    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
    sampling_rate = feature_extractor.sampling_rate
    print(sampling_rate)
    dataset = dataset_test.map(remove_special_characters)
    # resample audio
    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))

    processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
    model = Wav2Vec2ForCTC.from_pretrained(args.model_id) 
    # processor = AutoProcessor.from_pretrained(args.model_id)
    # model = AutoModelForCTC.from_pretrained(args.model_id) 
    model.to("cuda")
    # 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"] = batch["sentence"]
    #     return batch

    def evaluate(batch):
        inputs = processor(batch["audio"]["array"], return_tensors="pt",sampling_rate=sampling_rate, padding=True)
        with torch.no_grad():
            logits = model(inputs.input_values.to("cuda")).logits
        # pred_ids = torch.argmax(logits, dim=-1)
        # batch["prediction"] = processor.batch_decode(pred_ids)
        batch["prediction"] = processor.batch_decode(logits.cpu().numpy()).text
        batch["target"] =batch["sentence"]

        return batch

    result = dataset.map(evaluate, 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."
    )
    args = parser.parse_args()

    main(args)