dreamvoice/train_utils/src/dataset/dreamvc.py

import pandas as pd
import os
import random
import ast
import numpy as np
import torch
from einops import repeat, rearrange
import librosa

from torch.utils.data import Dataset
import torchaudio


class DreamData(Dataset):
    def __init__(self, data_dir, meta_dir, subset, prompt_dir,):
        self.datadir = data_dir
        meta = pd.read_csv(meta_dir)
        self.meta = meta[meta['subset'] == subset]
        self.subset = subset
        self.prompts = pd.read_csv(prompt_dir)

    def __getitem__(self, index):
        row = self.meta.iloc[index]

        # get spk
        spk_path = self.datadir + row['spk_path']
        spk = torch.load(spk_path, map_location='cpu').squeeze(0)

        speaker = row['speaker']

        # get prompt
        prompt = self.prompts[self.prompts['speaker_id'] == str(speaker)].sample(1)['prompt'].iloc[0]
        return spk, prompt

    def __len__(self):
        return len(self.meta)