流行天后孫燕姿的音色固然是極好的，但是目前全網都是她的聲音復刻，聽多了難免會有些審美疲勞，在網路上檢索了一圈，還沒有發現民謠歌手的音色模型，人就是這樣，得不到的永遠在騷動，本次我們自己構建訓練集，來打造自己的音色模型，讓民謠女神來唱流行歌曲，要多帶勁就有多帶勁。
構建訓練集
訓練集是指用於訓練神經網路模型的資料集合。這個資料集通常由大量的輸入和對應的輸出組成，神經網路模型透過學習輸入和輸出之間的關係來進行訓練，並且在訓練過程中調整模型的引數以最小化誤差。
通俗地講，如果我們想要訓練民謠歌手葉蓓的音色模型，就需要將她的歌曲作為輸入引數，也就是訓練集，訓練集的作用是為模型提供學習的材料，使其能夠從輸入資料中學習到正確的輸出。透過反覆迭代訓練集，神經網路模型可以不斷地最佳化自身，提高其對輸入資料的預測能力。
沒錯，so-vits庫底層就是神經網路架構，而訓練音色模型庫，本質上解決的是預測問題，關於神經網路架構，請移步：人工智慧機器學習底層原理剖析,人造神經元,您一定能看懂,通俗解釋把AI“黑話”轉化為“白話文”，這裡不再贅述。
選擇訓練集樣本時，最好選擇具有歌手音色“特質”的歌曲，為什麼全網都是孫燕姿？只是因為她的音色辨識度太高，模型可以從輸入資料中更容易地學習到正確的輸出。
此外，訓練集資料貴精不貴多，特徵權重比較高的清晰樣本，在訓練效果要比低質量樣本要好，比如歌手“翻唱”的一些歌曲，或者使用非常規唱法的歌曲，這類樣本雖然也具備一些歌手的音色特徵，但對於模型訓練來說，實際上起到是反作用，這是需要注意的事情。
這裡選擇葉蓓早期專輯《幸福深處》中的六首歌：
通常來說，訓練集的數量越多，模型的效能就越好，但是在實踐中，需要根據實際情況進行權衡和選擇。
在深度學習中，通常需要大量的資料才能訓練出高效能的模型。例如，在計算機視覺任務中，需要大量的影像資料來訓練卷積神經網路模型。但是，在其他一些任務中，如語音識別和自然語言處理，相對較少的資料量也可以訓練出高效能的模型。
通常，需要確保訓練集中包含充足、多樣的樣本，以覆蓋所有可能的輸入情況。此外，訓練集中需要包含足夠的正樣本和負樣本，以保證模型的分類效能。
除了數量之外，訓練集的質量也非常重要。需要確保訓練集中不存在偏差和噪聲，同時需要進行資料清洗和資料增強等預處理操作，以提高訓練集的質量和多樣性。
總的來說，訓練集的數量要求需要根據具體問題進行調整，需要考慮問題的複雜性、資料的多樣性、模型的複雜度和訓練演算法的效率等因素。在實踐中，需要進行實驗和驗證，找到最適合問題的訓練集規模。
綜上，考慮到筆者的電腦配置以及訓練時間成本，訓練集相對較小，其他朋友可以根據自己的情況豐儉由己地進行調整。
訓練集資料清洗
準備好訓練集之後，我們需要對資料進行“清洗”，也就是去掉歌曲中的伴奏、停頓以及混音部分，只留下“清唱”的版本。
伴奏和人聲分離推薦使用spleeter庫：
pip3 install spleeter --user
接著執行命令，對訓練集歌曲進行分離操作：
spleeter separate -o d:/output/ -p spleeter:2stems d:/資料.mp3
這裡-o代表輸出目錄，-p代表選擇的分離模型，最後是要分離的素材。
首次執行會比較慢，因為spleeter會下載預訓練模型，體積在1.73g左右，執行完畢後，會在輸出目錄生成分離後的音軌檔案：
D:\歌曲製作\清唱 的目錄
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 個檔案 252,203,784 位元組
2 個目錄 449,446,780,928 可用位元組
關於spleeter更多的操作，請移步至：人工智慧AI庫Spleeter免費人聲和背景音樂分離實踐(Python3.10), 這裡不再贅述。
分離後的資料樣本還需要二次處理，因為分離後的音訊本身還會帶有一些輕微的背景音和混音，這裡推薦使用noisereduce庫：
pip3 install noisereduce,soundfile
隨後進行降噪處理：
import noisereduce as nr
import soundfile as sf
# 讀入音訊檔案
data, rate = sf.read("audio_file.wav")
# 獲取噪聲樣本
noisy_part = data[10000:15000]
# 估算噪聲
noise = nr.estimate_noise(noisy_part, rate)
# 應用降噪演算法
reduced_noise = nr.reduce_noise(audio_clip=data, noise_clip=noise, verbose=False)
# 將結果寫入檔案
sf.write("audio_file_denoised.wav", reduced_noise, rate)
先透過soundfile庫將歌曲檔案讀出來，然後獲取噪聲樣本並對其使用降噪演算法，最後寫入新檔案。
至此，資料清洗工作基本完成。
訓練集資料切分
深度學習過程中，計算機會把訓練資料讀入顯示卡的快取中，但如果訓練集資料過大，會導致記憶體溢位問題，也就是常說的“爆視訊記憶體”現象。
將資料集分成多個部分，每次只載入一個部分的資料進行訓練。這種方法可以減少記憶體使用，同時也可以實現並行處理，提高訓練效率。
這裡可以使用github.com/openvpi/audio-slicer庫：
git clone https://github.com/openvpi/audio-slicer.git
隨後編寫程式碼：
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.
該指令碼可以將所有降噪後的清唱樣本切成小樣本，方便訓練，電腦配置比較低的朋友，可以考慮將min_interval和max_sil_kept調的更高一些，這些會切的更碎，所謂“細細切做臊子”。
最後，六首歌被切成了140個小樣本：
D:\歌曲製作\slicer 的目錄
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 個檔案 183,026,452 位元組
至此，資料切分順利完成。
開始訓練
萬事俱備，只差訓練，首先配置so-vits-svc環境，請移步：AI天后,線上飆歌,人工智慧AI孫燕姿模型應用實踐，復刻《遙遠的歌》，原唱晴子(Python3.10)，囿於篇幅，這裡不再贅述。
隨後將切分後的資料集放在專案根目錄的dataset_raw/yebei資料夾，如果沒有yebei資料夾，請進行建立。
隨後構建訓練配置檔案：
{
"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
}
}
這裡epochs是指對整個訓練集進行一次完整的訓練。具體來說，每個epoch包含多個訓練步驟，每個訓練步驟會從訓練集中抽取一個小批次的資料進行訓練，並更新模型的引數。
需要調整的引數是batch_size，如果視訊記憶體不夠，需要往下調整，否則也會“爆視訊記憶體”，如果訓練過程中出現了下面這個錯誤：
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
那麼就說明視訊記憶體已經不夠用了。
最後，執行命令開始訓練：
python3 train.py -c configs/config.json -m 44k
終端會返回訓練過程：
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
這裡每一次Epoch系統都會返回損失函式等相關資訊，訓練好的模型存放在專案的logs/44k目錄下，模型的字尾名是.pth。
結語
一般情況下，訓練損失率低於50%，並且損失函式在訓練集和驗證集上都趨於穩定，則可以認為模型已經收斂。收斂的模型就可以為我們所用了，如何使用訓練好的模型，請移步：AI天后,線上飆歌,人工智慧AI孫燕姿模型應用實踐，復刻《遙遠的歌》，原唱晴子(Python3.10)。
最後，奉上民謠女神葉蓓的總訓練6400次的音色模型，與眾鄉親同饗：
pan.baidu.com/s/1m3VGc7RktaO5snHw6RPLjQ?pwd=pqkb
提取碼：pqkb