tools/multi_thread/extract_hubert_code.py

#!/usr/bin/env python3
# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import logging
import torch
from tqdm import tqdm
import numpy as np
import json

import torch.distributed as distr
import pathlib
from distributed import init_distributed_context

import logging
logger = logging.getLogger(__name__)
import os
import sys
import re
import glob

sys.path.insert(0,'/apdcephfs_nj7/share_303172353/ggyzhang/projects/textlesslib')

import torchaudio
from textless.data.speech_encoder import SpeechEncoder
import soundfile
import torchaudio.transforms as T

def resample_wav(wav_data,sr,target_sr):
    if sr!=target_sr:
        resampler = T.Resample(orig_freq=sr, new_freq=target_sr)
        wav_data = resampler(wav_data)
    return wav_data


def single_job(encoder, wav_fp, save_fp, device,sample_rate=16000):
    item_name = os.path.basename(wav_fp).split('.')[0]
    data, sr = torchaudio.load(wav_fp)
    data = resample_wav(data,sr,target_sr=sample_rate)
    encoded = encoder(data.to(device))
    units = encoded["units"].detach().cpu().numpy()
    np.save(save_fp,units)


def extract_speech_token(args, rank, world_size):
    all_data = []
    test_fp = '/apdcephfs_nj7/share_303172353/ggyzhang/projects/v2s/data/lrs3/test_data.json'
    train_fp = '/apdcephfs_nj7/share_303172353/ggyzhang/projects/v2s/data/lrs3/train_data.json'
    valid_fp = '/apdcephfs_nj7/share_303172353/ggyzhang/projects/v2s/data/lrs3/valid_data_ori.json'
    with open(train_fp,'r') as fp:
        cur_data = json.load(fp)
        all_data.extend(cur_data)
    with open(valid_fp,'r') as fp:
        cur_data = json.load(fp)
        all_data.extend(cur_data)
    with open(test_fp,'r') as fp:
        cur_data = json.load(fp)
        all_data.extend(cur_data)
    wavs = [item['wav_fn'] for item in all_data]
    # wavs = glob.glob(f'{args.wav_dir}/**/*.wav',recursive=True)

    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    dense_model_name = "hubert-base-ls960-layer-9"
    quantizer_name, vocab_size = "kmeans", 500

    encoder = SpeechEncoder.by_name(
        dense_model_name=dense_model_name,
        quantizer_model_name=quantizer_name,
        vocab_size=vocab_size,
        deduplicate=False,
        dense_model_fp = '/apdcephfs_nj7/share_303172353/ggyzhang/projects/textlesslib/ckpts/hubert_base_ls960.pt',
        quantizer_model_fp='/apdcephfs_nj7/share_303172353/ggyzhang/projects/textlesslib/ckpts/hubert_base_ls960_L9_km500.bin'
    ).to(device)

    print(len(wavs))
    for i in tqdm(range(rank, len(wavs), world_size)):
        wav_fp = wavs[i]
        item_name = os.path.basename(wav_fp).split('.')[0]
        new_fp = os.path.dirname(wav_fp).replace('LRS3','LRS3_hubert_token')
        os.makedirs(new_fp,exist_ok=True)
        save_path = f'{new_fp}/{item_name}.npy'
        if os.path.exists(save_path):
            continue
        try:
            single_job(encoder,wav_fp,f'{new_fp}/{item_name}.npy',device)
        except:
            print('error!!!!!!!!',wav_fp)

def main(args):
    context = init_distributed_context(args.distributed_port)
    logger.info(f"Distributed context {context}")

    n_gpus = torch.cuda.device_count()
    with torch.cuda.device(context.local_rank % n_gpus):
        extract_speech_token(args, context.rank, context.world_size)

    if context.world_size > 1:
        distr.barrier()

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--distributed_port", type=int, default=58564)
    args = parser.parse_args()

    main(args)