import argparse
import os
import torch
from espnet2.bin.s2t_inference_ctc import Speech2TextGreedySearch
import evaluate
from normalizer import data_utils
import time
from tqdm import tqdm

wer_metric = evaluate.load("wer")

def main(args):
    # Load model
    device = "cpu" if args.device == -1 else f"cuda:{args.device}"
    load_args = {
        "device": device,
        "dtype": "float32",
        "use_flash_attn": True,
        "lang_sym": "<eng>",
        "task_sym": "<asr>",
    }
    if os.path.exists(args.model_id):
        load_args["s2t_model_file"] = args.model_id
    else:
        load_args["model_tag"] = args.model_id

    model = Speech2TextGreedySearch.from_pretrained(**load_args)
    print(f"Model size: {sum(p.numel() for p in model.s2t_model.parameters()) / 1e9:.2f}B parameters")

    def benchmark(batch):
        # Load audio inputs
        audios = [audio["array"] for audio in batch["audio"]]
        batch["audio_length_s"] = [len(audio) / batch["audio"][0]["sampling_rate"] for audio in audios]
        minibatch_size = len(audios)
        batch["audio_filepath"] = data_utils.extract_audio_filepaths_from_batch(batch, minibatch_size)

        # Start timing
        start_time = time.time()

        # INFERENCE
        with torch.autocast(device, torch.bfloat16, enabled=True), torch.inference_mode():
            pred_text = model.batch_decode(
                audios,
                batch_size=args.batch_size,
                context_len_in_secs=4,
            )

        # End timing
        runtime = time.time() - start_time

        # normalize by minibatch size since we want the per-sample time
        batch["transcription_time_s"] = minibatch_size * [runtime / minibatch_size]

        # normalize transcriptions with English normalizer
        batch["predictions"] = [data_utils.normalizer(pred) for pred in pred_text]
        batch["references"] = batch["norm_text"]
        return batch

    if args.warmup_steps is not None:
        warmup_dataset = data_utils.load_data(args)
        warmup_dataset = data_utils.prepare_data(warmup_dataset)

        num_warmup_samples = args.warmup_steps * args.batch_size
        if args.streaming:
            warmup_dataset = warmup_dataset.take(num_warmup_samples)
        else:
            warmup_dataset = warmup_dataset.select(range(min(num_warmup_samples, len(warmup_dataset))))
        warmup_dataset = iter(warmup_dataset.map(benchmark, batch_size=args.batch_size, batched=True))

        for _ in tqdm(warmup_dataset, desc="Warming up..."):
            continue

    dataset = data_utils.load_data(args)
    dataset = data_utils.prepare_data(dataset)

    if args.max_eval_samples is not None and args.max_eval_samples > 0:
        print(f"Subsampling dataset to first {args.max_eval_samples} samples!")
        if args.streaming:
            dataset = dataset.take(args.max_eval_samples)
        else:
            dataset = dataset.select(range(min(args.max_eval_samples, len(dataset))))

    dataset = dataset.map(
        benchmark, batch_size=args.batch_size, batched=True, remove_columns=["audio"],
    )

    all_results = {
        "audio_length_s": [],
        "transcription_time_s": [],
        "predictions": [],
        "references": [],
        "audio_filepath": [],
    }
    result_iter = iter(dataset)
    for result in tqdm(result_iter, desc="Samples..."):
        for key in all_results:
            all_results[key].append(result[key])

    # Write manifest results (WER and RTFX)
    manifest_path = data_utils.write_manifest(
        all_results["references"],
        all_results["predictions"],
        args.model_id,
        args.dataset_path,
        args.dataset,
        args.split,
        audio_length=all_results["audio_length_s"],
        transcription_time=all_results["transcription_time_s"],
        audio_filepaths=all_results["audio_filepath"],
    )
    print("Results saved at path:", os.path.abspath(manifest_path))

    wer = wer_metric.compute(
        references=all_results["references"], predictions=all_results["predictions"]
    )
    wer = round(100 * wer, 2)
    rtfx = round(sum(all_results["audio_length_s"]) / sum(all_results["transcription_time_s"]), 2)
    print("WER:", wer, "%", "RTFx:", rtfx)


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

    parser.add_argument(
        "--model_id",
        type=str,
        required=True,
        help="Model identifier. Should be loadable with ESPnet",
    )
    parser.add_argument(
        "--dataset_path",
        type=str,
        default="hf-audio/open-asr-leaderboard",
        help="Dataset path. By default, it is `hf-audio/open-asr-leaderboard`",
    )
    parser.add_argument(
        "--dataset",
        type=str,
        required=True,
        help="Dataset name. *E.g.* `'librispeech_asr` for the LibriSpeech ASR dataset, or `'common_voice'` for Common Voice. The full list of dataset names "
        "can be found at `https://huggingface.co/datasets/hf-audio/open-asr-leaderboard`",
    )
    parser.add_argument(
        "--split",
        type=str,
        default="test",
        help="Split of the dataset. *E.g.* `'validation`' for the dev split, or `'test'` for the test split.",
    )
    parser.add_argument(
        "--device",
        type=int,
        default=-1,
        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
    )
    parser.add_argument(
        "--batch_size",
        type=int,
        default=1,
        help="Number of samples to go through each streamed batch.",
    )
    parser.add_argument(
        "--max_eval_samples",
        type=int,
        default=None,
        help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.",
    )
    parser.add_argument(
        "--streaming",
        action="store_true",
        help="Stream the dataset lazily over the network instead of downloading it in full before the evaluation. Off by default for reproducible benchmark timings.",
    )
    parser.add_argument(
        "--warmup_steps",
        type=int,
        default=10,
        help="Number of warm-up steps to run before launching the timed runs.",
    )
    args = parser.parse_args()

    main(args)