eval.py

import re
import argparse
import unicodedata
from typing import Dict
import torch
import torchaudio
from datasets import load_dataset, load_metric,  Audio, Dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM
import re

chars_to_ignore_regex = '[\é\！\，\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\’\—\–\·]'

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["sentence"], predictions=result["pred_strings"])
    cer_result = cer.compute(references=result["sentence"], predictions=result["pred_strings"])

    # 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["pred_strings"] + "\n")
                t.write(f"{i}" + "\n")
                t.write(batch["sentence"] + "\n")

            result.map(write_to_file, with_indices=True)

def load_data(dataset_id, language, split='test'):
    test_dataset = load_dataset(dataset_id, language, split=split, use_auth_token=True)
    test_dataset = test_dataset.cast_column("audio", Audio(sampling_rate=16_000))
    return test_dataset

def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).lower() + " "
    batch["sentence"] = re.sub('!', '', batch["sentence"]).lower() + " "
    batch["sentence"] = batch["sentence"].replace('\"',"").replace("&","").replace("'","").replace("(","").lower() + " "
    batch["sentence"] = batch["sentence"].replace('[',"").replace("]","").replace("\\","").replace("«","").replace("»","").replace(")","").lower() + " "
    batch["sentence"] = batch["sentence"].replace("  "," ").replace("  "," ").replace("  "," ").lower() + " "
    batch["speech"] = batch["audio"]["array"]
    return batch

def main(args):
    test_dataset = load_data(args.dataset, args.config, args.split)
    test_dataset = test_dataset.map(speech_file_to_array_fn)
    model_id = args.model_id
    
    def evaluate_with_lm(batch):
        inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
        with torch.no_grad():
            logits = model(**inputs.to('cuda')).logits
        int_result = processor.batch_decode(logits.cpu().numpy())
        batch["pred_strings"] =  int_result.text
        return batch

    def evaluate(batch):
        inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
        with torch.no_grad():
            logits = model(inputs.input_values.to('cuda')).logits
        pred_ids = torch.argmax(logits, dim=-1)
        batch["pred_strings"] = processor.batch_decode(pred_ids, skip_special_tokens=True)
        return batch

    if args.lm:
        processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_id,use_auth_token=True)
        model = Wav2Vec2ForCTC.from_pretrained(model_id,use_auth_token=True)
        model.to('cuda')
        result = test_dataset.map(evaluate_with_lm, batched=True, batch_size=4)
    else:
        processor = Wav2Vec2Processor.from_pretrained(model_id,use_auth_token=True)
        model = Wav2Vec2ForCTC.from_pretrained(model_id,use_auth_token=True)
        model.to("cuda")
        result = test_dataset.map(evaluate, batched=True, batch_size=4)

    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(
        "--lm", action='store_true', help="Using language model for evaluation or not."
    )
    args = parser.parse_args()

    main(args)