| import os.path |
| import json |
|
|
| import numpy as np |
| import pandas as pd |
| import torch |
| from datasets import Dataset |
| from matplotlib import pyplot as plt |
| from tqdm import tqdm |
|
|
| phoneme_path = 'dump/phoneme.npy' |
| flist_path = 'dump/train_files_gs.list' |
| phoneme_data = np.load(phoneme_path, allow_pickle=True).item() |
|
|
| all_paths = [line.strip() for line in open(flist_path).readlines()] |
| all_paths = set(all_paths) |
| phoneme_data = {k: v for k, v in tqdm(phoneme_data.items()) if k in all_paths} |
| print(len(phoneme_data)) |
| def generate_data(): |
| for k, v in phoneme_data.items(): |
| itemname = k |
| code_path = itemname.replace(".wav", ".code.pt") |
| if not os.path.exists(code_path): |
| continue |
| code = torch.load(code_path).squeeze(0) |
|
|
| dur = code.shape[0] / 10 |
| spk_emb_path = itemname.replace(".wav", ".spk.npy") |
| if not os.path.exists(spk_emb_path): |
| print( f'spk_emb_path {spk_emb_path} not exists for {itemname}\n\n\n\n\n\n\n\n\n') |
| continue |
| spk_emb = np.load(spk_emb_path) |
|
|
| if spk_emb.shape != (256, ): |
| print( f'spk_emb shape {spk_emb.shape} not correct for {itemname}\n\n\n\n\n\n\n\n\n') |
| if dur > 100: |
| print(f'dur {dur} too long for {itemname}\n\n\n\n\n\n\n\n') |
| yield {"item_name": itemname, "codes": code, 'phoneme': v, 'length': dur, 'spk_emb': spk_emb} |
|
|
| dataset = Dataset.from_generator(generate_data) |
| print(dataset) |
|
|
| test_samples = 10 |
|
|
| split_dataset = dataset.train_test_split(test_size=test_samples/len(dataset), shuffle=True) |
|
|
| split_dataset.save_to_disk('dump/phone_semantic_dataset') |
| print("average duration", np.mean(dataset['length'])) |
