Model Fusion Error on Ilaria RVC Hugging Face Space

[dusty iron]

I'm trying to merge two models on Ilaria's RVC Hugging Face Space but I kept getting an output saying, "Fail to merge the models. The model architectures are not the same." The RVC models I'm using to merge are both trained with pitch guidance using a dataset with a sample rate of 32k and they are both version 2. The only difference is for model A which has 200 epochs and 2000 steps, and model B has 920 epochs and 9200 steps.

[unreal lake]

they must have the same v2 and sample rate, and I dont think there's a way to alter the model's sample rate

[dusty iron]

Both models I'm trying to merge do have the same v2 and sample rate but I kept getting the same error.

I've specified it earlier in the description:
"The RVC models I'm using to merge are both trained with pitch guidance using a dataset with a sample rate of 32k and they are both version 2."

I meant to say they use the same dataset.

[unreal lake]

have you tried in original rvc and applio? there shouldnt be any gpu requirement for doing it unlike the training

also you can use model information tool to make sure

[dusty iron]

I tried using Mainline RVC but they don't seem to be able to merge voice models that have a sample rate of 32k. I'll probably try Applio.

[unreal lake]

perhaps corrupted, have you tried other 32k models?

[dusty iron]

Not particularly. I was retraining a voice model from earlier downloading both path files with which has best loss in "g/fm" (200 epochs) and "g/mel" (920 epochs). I'm not sure if that's possible. Both do have a sample rate of 32k. I was gonna use the No UI Colab notebook of Applio but it doesn't have that feature (same could be said about the Colab notebook for Mainline RVC). I'll try using the UI Colab notebook for Applio to see if I can merge these two models.

[unreal lake]

again do other models work or not?

[dusty iron]

Well, I tried two ordinary models. Both are version 2 models with a sample rate of 32k and I still got the same error: "Fail to merge the models. The model architectures are not the same." I assume people are unable to fuse models if that were the case.

Also, once again, this was done on Ilaria's RVC Hugging Face Space.

I'm also getting the same results with 40k sample rate models (possibly the same with 48k sample rate models) so it looks like you can't do model fusion on Ilaria's RVC Hugging Face Space.

[unreal lake]

sorry have just checked it out, in original RVC on mine there is no 32k option, but it works on 40k models I tested

while in ilaria RVC yea there's issue on it, even on my 32k models, @marsh imp could you check it out?

[dusty iron]

I've also tried using the Applio Colab notebook with UI but there is no feature to do model fusion as well.

Nevermind, I realized it's under "voice blender."

After using "voice blender," I was able to merge two voice models together so I'll use Applio for now on then.

[marsh imp]

I don't think the Ilaria RVC Zero HuggingFace space was made for that too

Just like how the UVR does not work in it

@chilly pier did u ever test if model fusion worked?

If not, it's better to specify that in the guide

Also bc only Inference should be running on ZeroGPU

[unreal lake]

yeah I have tried it works only on Applio

[marsh imp]

I think it's because the HF space was supposed just for inference, as in the code the ZeroGPU decoration is only for Inference

[dusty iron]

There was this announcement made in Dev Updates where Ilaria mentioned having the feature to do model fusion on her RVC Hugging Face Space, but it looks like you can't because I've tried it with other models of different sample rates -- I've made sure both models have the same sample rates before doing so -- and it wouldn't work. #📰│dev-updates message

[marsh imp]

Well no I just tested myself and Model Fusion doesn't work

And UVR runs on CPU, don't think it works either

Just tried 10 secs Rick roll and nah UVR doesn't work

[unreal lake]

at this moment, I'd recommend using MSST colab where there are many good model options and you can add more models by modifying code in the separation cell (either the original jarredou's or tweaked version by me: #1159290752195633273 message)

[marsh imp]

I was just saying that it doesn't work, I didn't actually need to do UVR

But that's good