Errors Running Cloud UVR on Colab

[void vigil]

Followed every step as instructed on the documentation but turned out running into errors while executing on the Separation page:

/content
python3: can't open file '/content/inference.py': [Errno 2] No such file or directory```
The folders are set up, the ``input.wav`` includes my input file.

[buoyant juniper]

which colab notebook? I'd recommend this anyway https://github.com/jarredou/Music-Source-Separation-Training-Colab-Inference

[void vigil]

I used the Music_Source_Separation_Universal_Colab.ipynb here https://colab.research.google.com/github/Eddycrack864/Music-Source-Separation-Universal-Colab/blob/main/Music_Source_Separation_Universal_Colab.ipynb

[buoyant juniper]

the link I gave above is more updated and includes new models, see the metric comparison below (not all included in the colab notebook)

also you can try the tweaked version by me: https://colab.research.google.com/drive/1IC6Q1hLF55_tK6mhky0SWYKGVF9T5WsY?usp=drive_link

[void vigil]

Which model is the alternative to MDX23C 8KFFT InstVoc HQ 2? MelBand Roformer Kim | FT (by ZFTurbo & Bas Curtiz)?

[mystic canyon]

Ayo? @void vigil level 1 !!!

[buoyant juniper]

MDX23C HQ 2 is discontinued, only the HQ 1 is included in UVR and the MSST inference, and it surely has worse metric than the mentioned roformer models above

[void vigil]

The doc seems to still include them: https://docs.ai-hub.wtf/rvc/resources/vocal-isolation/#cloud-uvr

but thanks for the help. I will test them out

Do I usually need to touch/adjust these settings based on different audio files? I am talking about your tweaked version.

[buoyant juniper]

that's optimal settings I'm used to, so leave it if not sure

[void vigil]

posdata, dim_t and chunk size are the same and jarredou confirmed me dis

select the highest chunk size, leave dim_t as deafult, and use overlap 16

[buoyant juniper]

too bad the dim_t option doesnt work and always output identical results (invert the one with another and it will be complete -inf dB silence), so I have removed it

[buoyant juniper]

wonder why overlap 16 instead of 8? though yea I noticed less bleed on 8 than 2 and also this measurement

[void vigil]

more overlap improves results

but takes longer to process ofc, more overlap means better results but it's not always like that

on the music separation colab, overlap 16 should be fine as 30 doesn't make too much difference

to, you know, improve results.

[buoyant juniper]

does overlap in the new updated UVR work same as MSS colab? haven't tried it cuz I can't afford downloading with my crappy mobile data

[void vigil]

i think so, a overlap of 16 should work fine in the SSS colab

personally i recommend to not use more than 16 overlap as process will take longer and it isnt worthy tbh

imma talking about that later, now i gotta sleep before its too late.

sorry for leaving dis kinda quick but i gotta sleep now lol

[buoyant juniper]

reminds me of MDX23C overlap, I forgor did you use overlap 8 or 16 on it?

[void vigil]

i use overlap 16

[buoyant juniper]

I see

[void vigil]

anyways goodnight

[buoyant juniper]

fyi also use batch_size = 4 and there won't be popping artifacts, have tried on unwa's models as well as other roformers and it is also used in kim's model config

[void vigil]

there is no batch size config in the MSS colab

[buoyant juniper]

check update on my tweaked MSST colab

[void vigil]

oh now i see it

batch size 4 gives no artifacts on mel roformer models?

[buoyant juniper]

yea no artifacts at all in the spectrogram

[void vigil]

but inference will be slower ig, i'll test it out this weekend if i can lol

will you someday port this to kaggle? :p

[buoyant juniper]

#1159290752195633273 message

[buoyant juniper]

the processing time in colab is not that terribly slow, it is close to the audio duration itself

[buoyant juniper]

could you help trying and comparing new models as in this list, including the ones I havent tried:

• SYHFT v4
• bleed suppressor

idw keep having fomo on these new model updates lol

https://cdn.discordapp.com/attachments/1159290752195633273/1308791386077790258/image.png?ex=673fe336&is=673e91b6&hm=be24b2fd2c13614a40b8d2a18a770aada67413ad9c33b49d51684939595ea6cc&

[void vigil]

I may try it with a random song and sent you in dms the outputs of each model, so you can tell me if there's any audible or spectrogram difference

of course, gdrive link because files may be heavy for discord and i don't have nitro

[void vigil]

tho are these available in your colab?

syhft v4 and bleed suppressor?

[buoyant juniper]

not yet

[void vigil]

oh ok then, sorry but I wouldn't test until these models are in colab

because i don't have a nvidia gpu

[buoyant juniper]

one of songs I have tried is 21 Guns
you can try on that song too

[void vigil]

as I said before I won't be able to test the models as my pc isn't powerful enough for uvr5 and I have an amd igpu

[buoyant juniper]

those models may be available on mvsep, or you can add them in the colab's separation cell containing the list

[void vigil]

ok then, testing them if I have enough free time