[FAQ] How do I continue the training my old models without fully restart the training

[ingrid-z]

I think this might help the newbies. I'm also not sure how to do it. 🧇

[ghost]

I think this might help the newbies. I also not sure how to do it. 🧇

Use the same name of the model and version of rvc like if you used v2 select v2 in the training tab. i think you can change the pitch algorithm but i'm not sure. is just that. also you might get better support in IAHispano (the discord server of applio)

[ingrid-z]

I've tried fully copied my model v1 folder to the model v2 folder just in case then set at much higher train epochs. It does work, but I got 4 lines debug info, maybe these are not important

DEBUG:matplotlib:matplotlib data path: C:\Users\Ingrid\Desktop\Applio-RVC-Fork\Applio-RVC-Fork\runtime\lib\site-packages\matplotlib\mpl-data
DEBUG:matplotlib:CONFIGDIR=C:\Users\Ingrid\.matplotlib
DEBUG:matplotlib:interactive is False
DEBUG:matplotlib:platform is win32

Use the same setting, you can skip first 2 steps(process and extract), jump into step 3 which is train. But I found my training time is double from 44s to 1:56 and I don't know why.

[ghost]

I've tried fully copied my model v1 folder to the model v2 folder just in case then set at much higher train period. It does work, but I got 4 lines debug info, maybe these are not important

DEBUG:matplotlib:matplotlib data path: C:\Users\Ingrid\Desktop\Applio-RVC-Fork\Applio-RVC-Fork\runtime\lib\site-packages\matplotlib\mpl-data
DEBUG:matplotlib:CONFIGDIR=C:\Users\Ingrid\.matplotlib
DEBUG:matplotlib:interactive is False
DEBUG:matplotlib:platform is win32

Use the same setting, you can skip first 2 steps(process and extract), jump into step 3 which is train. But I found my training time is double from 44s to 1:56 and I don't know why.

You can't trin from rvc 1 to 2 i think

[ingrid-z]

You can't trin from rvc 1 to 2 i think

I'm afraid you're misunderstanding my word. I said I create a brand new folder and copied all my old file into new folder, the train setting was totally the same. Besides I increased the train epochs.

[ghost]

You can't trin from rvc 1 to 2 i think

I'm afraid you're misunderstanding my word. I said I create a brand new folder and copied all my old file into new folder, the train setting was totally the same. Besides I increased the train epochs.

uh thats weird

[ghost]

Anyways this should be fixed with the latest commits

[ingrid-z]

I found a way to continue the old models training. Step1. Create the new model name under /logs and copy the old model from /logs/model name that you want to continue training. Step2. Dataset must come from the old model you want to to continue training. or you can clone dataset into new folder(wasting space lmao once dataset processed it'll never be changed later.) Step3. Make sure your setting is match with the old model setting. ig pitch algorithm, hop length Step4. Do NOT do data process neither extract feature, because those steps are already done in old model. Step5. Change Train Epochs to higher than original old model. (opitonal: change batch size if you have a high end GPU) Done

Somehow there is a bug. if you stop the traing during training time. and raise the train epochs again that being said you've changed epochs twice in this session. then start the training. it will be very very slow in analyzing phase seems likes stuck in the phase forever.

[ghost]

Thanks for that information. im going to leave it here but in spanish

Paso 1. Ingrese el nuevo nombre del modelo y copie el modelo anterior de /log/nombre_del_modelo con el que desea continuar entrenando. Paso 2. El dataset debe provenir del modelo anterior que desea continuar entrenando. Paso 3. Asegúrese de que su configuración coincida con la configuración del modelo anterior. algoritmo de tono, hop length Etapa 4. NO procese datos ni extraiga caracterisiticas, porque esos pasos ya se realizan en el entrenamiento anterior. Paso 5. Cambie los epochs o epocas a unas superiores que el modelo antiguo. (opcional: cambie el tamaño del lote (batch size) si tiene una GPU de gama alta) Hecho

De alguna manera hay un error. si detienes el entrenamiento durante el tiempo de entrenamiento. y vuelva a subir las épocas del tren. Dicho esto, ha cambiado dos veces las épocas en esta sesión. Luego comienza el entrenamiento. Será muy, muy lento en la fase de análisis. Parece estar atrapado en la fase para siempre.