流行天后孙燕姿的音色固然是极好的，可是现在全网都是她的声响复刻，听多了难免会有些审美疲劳，在网络上检索了一圈，还没有发现歌谣歌手的音色模型，人便是这样，得不到的永远在骚乱，本次咱们自己构建练习集，来打造自己的音色模型，让歌谣女神来唱流行歌曲，要多带劲就有多带劲。
构建练习集
练习集是指用于练习神经网络模型的数据调集。这个数据集通常由很多的输入和对应的输出组成，神经网络模型通过学习输入和输出之间的关系来进行练习，而且在练习过程中调整模型的参数以最小化差错。
通俗地讲，假如咱们想要练习歌谣歌手叶蓓的音色模型，就需求将她的歌曲作为输入参数，也便是练习集，练习集的作用是为模型供给学习的资料，使其能够从输入数据中学习到正确的输出。通过反复迭代练习集，神经网络模型能够不断地优化自身，进步其对输入数据的预测才干。
没错，so-vits库底层便是神经网络架构，而练习音色模型库，本质上处理的是预测问题，关于神经网络架构，请移步：人工智能机器学习底层原理剖析,人工神经元,您必定能看懂,通俗解释把AI“黑话”转化为“白话文”，这儿不再赘述。
挑选练习集样本时，最好挑选具有歌手音色“特质”的歌曲，为什么全网都是孙燕姿？仅仅因为她的音色辨识度太高，模型能够从输入数据中更容易地学习到正确的输出。
此外，练习集数据贵精不贵多，特征权重比较高的清晰样本，在练习作用要比低质量样本要好，比方歌手“翻唱”的一些歌曲，或者运用非常规唱法的歌曲，这类样本尽管也具有一些歌手的音色特征，但关于模型练习来说，实际上起到是反作用，这是需求留意的事情。
这儿挑选叶蓓早期专辑《美好深处》中的六首歌：
通常来说，练习集的数量越多，模型的性能就越好，可是在实践中，需求依据实际情况进行权衡和挑选。
在深度学习中，通常需求很多的数据才干练习出高性能的模型。例如，在计算机视觉任务中，需求很多的图像数据来练习卷积神经网络模型。可是，在其他一些任务中，如语音辨认和自然语言处理，相对较少的数据量也能够练习出高性能的模型。
通常，需求确保练习会集包括足够、多样的样本，以覆盖一切可能的输入情况。此外，练习会集需求包括足够的正样本和负样本，以保证模型的分类性能。
除了数量之外，练习集的质量也非常重要。需求确保练习会集不存在偏差和噪声，一起需求进行数据清洗和数据增强等预处理操作，以进步练习集的质量和多样性。
总的来说，练习集的数量要求需求依据具体问题进行调整，需求考虑问题的复杂性、数据的多样性、模型的复杂度和练习算法的效率等要素。在实践中，需求进行试验和验证，找到最适合问题的练习集规模。
综上，考虑到笔者的电脑装备以及练习时刻本钱，练习集相对较小，其他朋友能够依据自己的情况丰俭由己地进行调整。
练习集数据清洗
准备好练习集之后，咱们需求对数据进行“清洗”，也便是去掉歌曲中的配乐、停顿以及混音部分，只留下“清唱”的版本。
配乐和人声别离推荐运用spleeter库：
pip3 install spleeter --user
接着运转命令，对练习集歌曲进行别离操作：
spleeter separate -o d:/output/ -p spleeter:2stems d:/数据.mp3
这儿-o代表输出目录，-p代表挑选的别离模型，最终是要别离的资料。
初次运转会比较慢，因为spleeter会下载预练习模型，体积在1.73g左右，运转完毕后，会在输出目录生成别离后的音轨文件：
D:\歌曲制造\清唱 的目录
2023/05/11 15:38 <DIR> .
2023/05/11 13:45 <DIR> ..
2023/05/11 13:40 39,651,884 1_1_01. wxs.wav
2023/05/11 15:34 46,103,084 1_1_02. qad_(Vocals)_(Vocals).wav
2023/05/11 15:35 43,802,924 1_1_03. hs_(Vocals)_(Vocals).wav
2023/05/11 15:36 39,054,764 1_1_04. hope_(Vocals)_(Vocals).wav
2023/05/11 15:36 32,849,324 1_1_05. kamen_(Vocals)_(Vocals).wav
2023/05/11 15:37 50,741,804 1_1_06. ctrl_(Vocals)_(Vocals).wav
6 个文件 252,203,784 字节
2 个目录 449,446,780,928 可用字节
关于spleeter更多的操作，请移步至：人工智能AI库Spleeter免费人声和背景音乐别离实践(Python3.10), 这儿不再赘述。
别离后的数据样本还需求二次处理，因为别离后的音频自身还会带有一些细微的背景音和混音，这儿推荐运用noisereduce库：
pip3 install noisereduce,soundfile
随后进行降噪处理：
import noisereduce as nr
import soundfile as sf
# 读入音频文件
data, rate = sf.read("audio_file.wav")
# 获取噪声样本
noisy_part = data[10000:15000]
# 估算噪声
noise = nr.estimate_noise(noisy_part, rate)
# 运用降噪算法
reduced_noise = nr.reduce_noise(audio_clip=data, noise_clip=noise, verbose=False)
# 将结果写入文件
sf.write("audio_file_denoised.wav", reduced_noise, rate)
先通过soundfile库将歌曲文件读出来，然后获取噪声样本并对其运用降噪算法，最终写入新文件。
至此，数据清洗作业基本完成。
练习集数据切分
深度学习过程中，计算机会把练习数据读入显卡的缓存中，但假如练习集数据过大，会导致内存溢出问题，也便是常说的“爆显存”现象。
将数据集分红多个部分，每次只载入一个部分的数据进行练习。这种办法能够削减内存运用，一起也能够完成并行处理，进步练习效率。
这儿能够运用github.com/openvpi/audio-slicer库：
git clone https://github.com/openvpi/audio-slicer.git
随后编写代码：
import librosa # Optional. Use any library you like to read audio files.
import soundfile # Optional. Use any library you like to write audio files.
from slicer2 import Slicer
audio, sr = librosa.load('example.wav', sr=None, mono=False) # Load an audio file with librosa.
slicer = Slicer(
sr=sr,
threshold=-40,
min_length=5000,
min_interval=300,
hop_size=10,
max_sil_kept=500
)
chunks = slicer.slice(audio)
for i, chunk in enumerate(chunks):
if len(chunk.shape) > 1:
chunk = chunk.T # Swap axes if the audio is stereo.
soundfile.write(f'clips/example_{i}.wav', chunk, sr) # Save sliced audio files with soundfile.
该脚本能够将一切降噪后的清唱样本切成小样本，便利练习，电脑装备比较低的朋友，能够考虑将min_interval和max_sil_kept调的更高一些，这些会切的更碎，所谓“细细切做臊子”。
最终，六首歌被切成了140个小样本：
D:\歌曲制造\slicer 的目录
2023/05/11 15:45 <DIR> .
2023/05/11 13:45 <DIR> ..
2023/05/11 15:45 873,224 1_1_01. wxs_0.wav
2023/05/11 15:45 934,964 1_1_01. wxs_1.wav
2023/05/11 15:45 1,039,040 1_1_01. wxs_10.wav
2023/05/11 15:45 1,391,840 1_1_01. wxs_11.wav
2023/05/11 15:45 2,272,076 1_1_01. wxs_12.wav
2023/05/11 15:45 2,637,224 1_1_01. wxs_13.wav
2023/05/11 15:45 1,476,512 1_1_01. wxs_14.wav
2023/05/11 15:45 1,044,332 1_1_01. wxs_15.wav
2023/05/11 15:45 1,809,908 1_1_01. wxs_16.wav
2023/05/11 15:45 887,336 1_1_01. wxs_17.wav
2023/05/11 15:45 952,604 1_1_01. wxs_18.wav
2023/05/11 15:45 989,648 1_1_01. wxs_19.wav
2023/05/11 15:45 957,896 1_1_01. wxs_2.wav
2023/05/11 15:45 231,128 1_1_01. wxs_20.wav
2023/05/11 15:45 1,337,156 1_1_01. wxs_3.wav
2023/05/11 15:45 1,308,932 1_1_01. wxs_4.wav
2023/05/11 15:45 1,035,512 1_1_01. wxs_5.wav
2023/05/11 15:45 2,388,500 1_1_01. wxs_6.wav
2023/05/11 15:45 2,952,980 1_1_01. wxs_7.wav
2023/05/11 15:45 929,672 1_1_01. wxs_8.wav
2023/05/11 15:45 878,516 1_1_01. wxs_9.wav
2023/05/11 15:45 963,188 1_1_02. qad_(Vocals)_(Vocals)_0.wav
2023/05/11 15:45 901,448 1_1_02. qad_(Vocals)_(Vocals)_1.wav
2023/05/11 15:45 1,411,244 1_1_02. qad_(Vocals)_(Vocals)_10.wav
2023/05/11 15:45 2,070,980 1_1_02. qad_(Vocals)_(Vocals)_11.wav
2023/05/11 15:45 2,898,296 1_1_02. qad_(Vocals)_(Vocals)_12.wav
2023/05/11 15:45 885,572 1_1_02. qad_(Vocals)_(Vocals)_13.wav
2023/05/11 15:45 841,472 1_1_02. qad_(Vocals)_(Vocals)_14.wav
2023/05/11 15:45 876,752 1_1_02. qad_(Vocals)_(Vocals)_15.wav
2023/05/11 15:45 1,091,960 1_1_02. qad_(Vocals)_(Vocals)_16.wav
2023/05/11 15:45 1,188,980 1_1_02. qad_(Vocals)_(Vocals)_17.wav
2023/05/11 15:45 1,446,524 1_1_02. qad_(Vocals)_(Vocals)_18.wav
2023/05/11 15:45 924,380 1_1_02. qad_(Vocals)_(Vocals)_19.wav
2023/05/11 15:45 255,824 1_1_02. qad_(Vocals)_(Vocals)_2.wav
2023/05/11 15:45 1,718,180 1_1_02. qad_(Vocals)_(Vocals)_20.wav
2023/05/11 15:45 2,070,980 1_1_02. qad_(Vocals)_(Vocals)_21.wav
2023/05/11 15:45 2,827,736 1_1_02. qad_(Vocals)_(Vocals)_22.wav
2023/05/11 15:45 862,640 1_1_02. qad_(Vocals)_(Vocals)_23.wav
2023/05/11 15:45 1,628,216 1_1_02. qad_(Vocals)_(Vocals)_24.wav
2023/05/11 15:45 1,626,452 1_1_02. qad_(Vocals)_(Vocals)_25.wav
2023/05/11 15:45 1,499,444 1_1_02. qad_(Vocals)_(Vocals)_26.wav
2023/05/11 15:45 1,303,640 1_1_02. qad_(Vocals)_(Vocals)_27.wav
2023/05/11 15:45 998,468 1_1_02. qad_(Vocals)_(Vocals)_28.wav
2023/05/11 15:45 781,496 1_1_02. qad_(Vocals)_(Vocals)_3.wav
2023/05/11 15:45 1,368,908 1_1_02. qad_(Vocals)_(Vocals)_4.wav
2023/05/11 15:45 892,628 1_1_02. qad_(Vocals)_(Vocals)_5.wav
2023/05/11 15:45 1,386,548 1_1_02. qad_(Vocals)_(Vocals)_6.wav
2023/05/11 15:45 883,808 1_1_02. qad_(Vocals)_(Vocals)_7.wav
2023/05/11 15:45 952,604 1_1_02. qad_(Vocals)_(Vocals)_8.wav
2023/05/11 15:45 1,303,640 1_1_02. qad_(Vocals)_(Vocals)_9.wav
2023/05/11 15:45 1,354,796 1_1_03. hs_(Vocals)_(Vocals)_0.wav
2023/05/11 15:45 1,344,212 1_1_03. hs_(Vocals)_(Vocals)_1.wav
2023/05/11 15:45 1,305,404 1_1_03. hs_(Vocals)_(Vocals)_10.wav
2023/05/11 15:45 1,291,292 1_1_03. hs_(Vocals)_(Vocals)_11.wav
2023/05/11 15:45 1,338,920 1_1_03. hs_(Vocals)_(Vocals)_12.wav
2023/05/11 15:45 1,093,724 1_1_03. hs_(Vocals)_(Vocals)_13.wav
2023/05/11 15:45 1,375,964 1_1_03. hs_(Vocals)_(Vocals)_14.wav
2023/05/11 15:45 1,409,480 1_1_03. hs_(Vocals)_(Vocals)_15.wav
2023/05/11 15:45 1,481,804 1_1_03. hs_(Vocals)_(Vocals)_16.wav
2023/05/11 15:45 2,247,380 1_1_03. hs_(Vocals)_(Vocals)_17.wav
2023/05/11 15:45 1,312,460 1_1_03. hs_(Vocals)_(Vocals)_18.wav
2023/05/11 15:45 1,428,884 1_1_03. hs_(Vocals)_(Vocals)_19.wav
2023/05/11 15:45 1,051,388 1_1_03. hs_(Vocals)_(Vocals)_2.wav
2023/05/11 15:45 1,377,728 1_1_03. hs_(Vocals)_(Vocals)_20.wav
2023/05/11 15:45 1,485,332 1_1_03. hs_(Vocals)_(Vocals)_21.wav
2023/05/11 15:45 897,920 1_1_03. hs_(Vocals)_(Vocals)_22.wav
2023/05/11 15:45 1,591,172 1_1_03. hs_(Vocals)_(Vocals)_23.wav
2023/05/11 15:45 920,852 1_1_03. hs_(Vocals)_(Vocals)_24.wav
2023/05/11 15:45 1,046,096 1_1_03. hs_(Vocals)_(Vocals)_25.wav
2023/05/11 15:45 730,340 1_1_03. hs_(Vocals)_(Vocals)_26.wav
2023/05/11 15:45 1,383,020 1_1_03. hs_(Vocals)_(Vocals)_3.wav
2023/05/11 15:45 1,188,980 1_1_03. hs_(Vocals)_(Vocals)_4.wav
2023/05/11 15:45 1,003,760 1_1_03. hs_(Vocals)_(Vocals)_5.wav
2023/05/11 15:45 1,243,664 1_1_03. hs_(Vocals)_(Vocals)_6.wav
2023/05/11 15:45 845,000 1_1_03. hs_(Vocals)_(Vocals)_7.wav
2023/05/11 15:45 892,628 1_1_03. hs_(Vocals)_(Vocals)_8.wav
2023/05/11 15:45 539,828 1_1_03. hs_(Vocals)_(Vocals)_9.wav
2023/05/11 15:45 725,048 1_1_04. hope_(Vocals)_(Vocals)_0.wav
2023/05/11 15:45 1,023,164 1_1_04. hope_(Vocals)_(Vocals)_1.wav
2023/05/11 15:45 202,904 1_1_04. hope_(Vocals)_(Vocals)_10.wav
2023/05/11 15:45 659,780 1_1_04. hope_(Vocals)_(Vocals)_11.wav
2023/05/11 15:45 1,017,872 1_1_04. hope_(Vocals)_(Vocals)_12.wav
2023/05/11 15:45 1,495,916 1_1_04. hope_(Vocals)_(Vocals)_13.wav
2023/05/11 15:45 1,665,260 1_1_04. hope_(Vocals)_(Vocals)_14.wav
2023/05/11 15:45 675,656 1_1_04. hope_(Vocals)_(Vocals)_15.wav
2023/05/11 15:45 1,187,216 1_1_04. hope_(Vocals)_(Vocals)_16.wav
2023/05/11 15:45 1,201,328 1_1_04. hope_(Vocals)_(Vocals)_17.wav
2023/05/11 15:45 1,368,908 1_1_04. hope_(Vocals)_(Vocals)_18.wav
2023/05/11 15:45 1,462,400 1_1_04. hope_(Vocals)_(Vocals)_19.wav
2023/05/11 15:45 963,188 1_1_04. hope_(Vocals)_(Vocals)_2.wav
2023/05/11 15:45 1,121,948 1_1_04. hope_(Vocals)_(Vocals)_20.wav
2023/05/11 15:45 165,860 1_1_04. hope_(Vocals)_(Vocals)_21.wav
2023/05/11 15:45 1,116,656 1_1_04. hope_(Vocals)_(Vocals)_3.wav
2023/05/11 15:45 622,736 1_1_04. hope_(Vocals)_(Vocals)_4.wav
2023/05/11 15:45 1,349,504 1_1_04. hope_(Vocals)_(Vocals)_5.wav
2023/05/11 15:45 984,356 1_1_04. hope_(Vocals)_(Vocals)_6.wav
2023/05/11 15:45 2,104,496 1_1_04. hope_(Vocals)_(Vocals)_7.wav
2023/05/11 15:45 1,762,280 1_1_04. hope_(Vocals)_(Vocals)_8.wav
2023/05/11 15:45 1,116,656 1_1_04. hope_(Vocals)_(Vocals)_9.wav
2023/05/11 15:45 1,114,892 1_1_05. kamen_(Vocals)_(Vocals)_0.wav
2023/05/11 15:45 874,988 1_1_05. kamen_(Vocals)_(Vocals)_1.wav
2023/05/11 15:45 1,400,660 1_1_05. kamen_(Vocals)_(Vocals)_10.wav
2023/05/11 15:45 943,784 1_1_05. kamen_(Vocals)_(Vocals)_11.wav
2023/05/11 15:45 1,351,268 1_1_05. kamen_(Vocals)_(Vocals)_12.wav
2023/05/11 15:45 1,476,512 1_1_05. kamen_(Vocals)_(Vocals)_13.wav
2023/05/11 15:45 933,200 1_1_05. kamen_(Vocals)_(Vocals)_14.wav
2023/05/11 15:45 1,388,312 1_1_05. kamen_(Vocals)_(Vocals)_15.wav
2023/05/11 15:45 1,012,580 1_1_05. kamen_(Vocals)_(Vocals)_16.wav
2023/05/11 15:45 1,365,380 1_1_05. kamen_(Vocals)_(Vocals)_17.wav
2023/05/11 15:45 1,614,104 1_1_05. kamen_(Vocals)_(Vocals)_18.wav
2023/05/11 15:45 1,582,352 1_1_05. kamen_(Vocals)_(Vocals)_19.wav
2023/05/11 15:45 949,076 1_1_05. kamen_(Vocals)_(Vocals)_2.wav
2023/05/11 15:45 1,402,424 1_1_05. kamen_(Vocals)_(Vocals)_20.wav
2023/05/11 15:45 1,268,360 1_1_05. kamen_(Vocals)_(Vocals)_21.wav
2023/05/11 15:45 1,016,108 1_1_05. kamen_(Vocals)_(Vocals)_22.wav
2023/05/11 15:45 1,065,500 1_1_05. kamen_(Vocals)_(Vocals)_3.wav
2023/05/11 15:45 874,988 1_1_05. kamen_(Vocals)_(Vocals)_4.wav
2023/05/11 15:45 954,368 1_1_05. kamen_(Vocals)_(Vocals)_5.wav
2023/05/11 15:45 1,049,624 1_1_05. kamen_(Vocals)_(Vocals)_6.wav
2023/05/11 15:45 878,516 1_1_05. kamen_(Vocals)_(Vocals)_7.wav
2023/05/11 15:45 1,019,636 1_1_05. kamen_(Vocals)_(Vocals)_8.wav
2023/05/11 15:45 1,383,020 1_1_05. kamen_(Vocals)_(Vocals)_9.wav
2023/05/11 15:45 1,005,524 1_1_06. ctrl_(Vocals)_(Vocals)_0.wav
2023/05/11 15:45 1,090,196 1_1_06. ctrl_(Vocals)_(Vocals)_1.wav
2023/05/11 15:45 84,716 1_1_06. ctrl_(Vocals)_(Vocals)_10.wav
2023/05/11 15:45 857,348 1_1_06. ctrl_(Vocals)_(Vocals)_11.wav
2023/05/11 15:45 991,412 1_1_06. ctrl_(Vocals)_(Vocals)_12.wav
2023/05/11 15:45 1,121,948 1_1_06. ctrl_(Vocals)_(Vocals)_13.wav
2023/05/11 15:45 931,436 1_1_06. ctrl_(Vocals)_(Vocals)_14.wav
2023/05/11 15:45 3,129,380 1_1_06. ctrl_(Vocals)_(Vocals)_15.wav
2023/05/11 15:45 6,202,268 1_1_06. ctrl_(Vocals)_(Vocals)_16.wav
2023/05/11 15:45 1,457,108 1_1_06. ctrl_(Vocals)_(Vocals)_17.wav
2023/05/11 15:45 1,046,096 1_1_06. ctrl_(Vocals)_(Vocals)_2.wav
2023/05/11 15:45 956,132 1_1_06. ctrl_(Vocals)_(Vocals)_3.wav
2023/05/11 15:45 1,286,000 1_1_06. ctrl_(Vocals)_(Vocals)_4.wav
2023/05/11 15:45 804,428 1_1_06. ctrl_(Vocals)_(Vocals)_5.wav
2023/05/11 15:45 1,337,156 1_1_06. ctrl_(Vocals)_(Vocals)_6.wav
2023/05/11 15:45 1,372,436 1_1_06. ctrl_(Vocals)_(Vocals)_7.wav
2023/05/11 15:45 2,954,744 1_1_06. ctrl_(Vocals)_(Vocals)_8.wav
2023/05/11 15:45 6,112,304 1_1_06. ctrl_(Vocals)_(Vocals)_9.wav
140 个文件 183,026,452 字节
至此，数据切分顺利完成。
开端练习
万事俱备，只差练习，首先装备so-vits-svc环境，请移步：AI天后,在线飙歌,人工智能AI孙燕姿模型运用实践，复刻《悠远的歌》，原唱晴子(Python3.10)，囿于篇幅，这儿不再赘述。
随后将切分后的数据集放在项目根目录的dataset_raw/yebei文件夹，假如没有yebei文件夹，请进行创建。
随后构建练习装备文件：
{
"train": {
"log_interval": 200,
"eval_interval": 800,
"seed": 1234,
"epochs": 10000,
"learning_rate": 0.0001,
"betas": [
0.8,
0.99
],
"eps": 1e-09,
"batch_size": 6,
"fp16_run": false,
"lr_decay": 0.999875,
"segment_size": 10240,
"init_lr_ratio": 1,
"warmup_epochs": 0,
"c_mel": 45,
"c_kl": 1.0,
"use_sr": true,
"max_speclen": 512,
"port": "8001",
"keep_ckpts": 10,
"all_in_mem": false
},
"data": {
"training_files": "filelists/train.txt",
"validation_files": "filelists/val.txt",
"max_wav_value": 32768.0,
"sampling_rate": 44100,
"filter_length": 2048,
"hop_length": 512,
"win_length": 2048,
"n_mel_channels": 80,
"mel_fmin": 0.0,
"mel_fmax": 22050
},
"model": {
"inter_channels": 192,
"hidden_channels": 192,
"filter_channels": 768,
"n_heads": 2,
"n_layers": 6,
"kernel_size": 3,
"p_dropout": 0.1,
"resblock": "1",
"resblock_kernel_sizes": [
3,
7,
11
],
"resblock_dilation_sizes": [
[
1,
3,
5
],
[
1,
3,
5
],
[
1,
3,
5
]
],
"upsample_rates": [
8,
8,
2,
2,
2
],
"upsample_initial_channel": 512,
"upsample_kernel_sizes": [
16,
16,
4,
4,
4
],
"n_layers_q": 3,
"use_spectral_norm": false,
"gin_channels": 768,
"ssl_dim": 768,
"n_speakers": 1
},
"spk": {
"yebei": 0
}
}
这儿epochs是指对整个练习集进行一次完好的练习。具体来说，每个epoch包括多个练习过程，每个练习过程会从练习会集抽取一个小批量的数据进行练习，并更新模型的参数。
需求调整的参数是batch_size，假如显存不够，需求往下调整，否则也会“爆显存”，假如练习过程中呈现了下面这个过错：
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 8.00 GiB total capacity; 6.86 GiB already allocated; 0 bytes free; 7.25 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
那么就说明显存现已不够用了。
最终，运转命令开端练习：
python3 train.py -c configs/config.json -m 44k
终端会回来练习过程：
D:\work\so-vits-svc\workenv\lib\site-packages\torch\optim\lr_scheduler.py:139: UserWarning: Detected call of `lr_scheduler.step()` before `optimizer.step()`. In PyTorch 1.1.0 and later, you should call them in the opposite order: `optimizer.step()` before `lr_scheduler.step()`. Failure to do this will result in PyTorch skipping the first value of the learning rate schedule. See more details at https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
warnings.warn("Detected call of `lr_scheduler.step()` before `optimizer.step()`. "
D:\work\so-vits-svc\workenv\lib\site-packages\torch\functional.py:641: UserWarning: stft with return_complex=False is deprecated. In a future pytorch release, stft will return complex tensors for all inputs, and return_complex=False will raise an error.
Note: you can still call torch.view_as_real on the complex output to recover the old return format. (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\SpectralOps.cpp:867.)
return _VF.stft(input, n_fft, hop_length, win_length, window, # type: ignore[attr-defined]
INFO:torch.nn.parallel.distributed:Reducer buckets have been rebuilt in this iteration.
D:\work\so-vits-svc\workenv\lib\site-packages\torch\autograd\__init__.py:200: UserWarning: Grad strides do not match bucket view strides. This may indicate grad was not created according to the gradient layout contract, or that the param's strides changed since DDP was constructed. This is not an error, but may impair performance.
grad.sizes() = [32, 1, 4], strides() = [4, 1, 1]
bucket_view.sizes() = [32, 1, 4], strides() = [4, 4, 1] (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\distributed\c10d\reducer.cpp:337.)
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
INFO:torch.nn.parallel.distributed:Reducer buckets have been rebuilt in this iteration.
INFO:44k:====> Epoch: 274, cost 39.02 s
INFO:44k:====> Epoch: 275, cost 17.47 s
INFO:44k:====> Epoch: 276, cost 17.74 s
INFO:44k:====> Epoch: 277, cost 17.43 s
INFO:44k:====> Epoch: 278, cost 17.59 s
INFO:44k:====> Epoch: 279, cost 17.82 s
INFO:44k:====> Epoch: 280, cost 17.64 s
INFO:44k:====> Epoch: 281, cost 17.63 s
INFO:44k:Train Epoch: 282 [65%]
INFO:44k:Losses: [1.8697402477264404, 3.029414415359497, 11.415563583374023, 23.37869644165039, 0.2702481746673584], step: 6600, lr: 9.637943809624507e-05, reference_loss: 39.963661193847656
这儿每一次Epoch系统都会回来丢失函数等相关信息，练习好的模型存放在项目的logs/44k目录下，模型的后缀名是.pth。
结语
一般情况下，练习丢失率低于50%，而且丢失函数在练习集和验证集上都趋于稳定，则能够认为模型现已收敛。收敛的模型就能够为咱们所用了，怎么运用练习好的模型，请移步：AI天后,在线飙歌,人工智能AI孙燕姿模型运用实践，复刻《悠远的歌》，原唱晴子(Python3.10)。
最终，奉上歌谣女神叶蓓的总练习6400次的音色模型，与众乡亲同飨：
pan.baidu.com/s/1m3VGc7RktaO5snHw6RPLjQ?pwd=pqkb
提取码：pqkb
