htdemucs model pretrained on only MUSDB18-HQ?

facebookresearch / demucs

Code for the paper Hybrid Spectrogram and Waveform Source Separation

MIT License

8.18k stars 1.03k forks source link

htdemucs model pretrained on only MUSDB18-HQ? #572

Open mrdaly opened 10 months ago

mrdaly commented 10 months ago

❓ Questions

Hi,

Are there any checkpoints for the Hybrid Transformer Demucs architecture that were only trained on the MUSDB18-HQ, without the additional internal data set mentioned in the paper?

ALEXander22568 commented 9 months ago

Hi,I have the same question with you ,I want to train a modle with new datasets but I don't sure whether I can make it

CarlGao4 commented 9 months ago

I'm doing this. I'll upload the model and checkpoints here once it's finished. I'm training the model with only MusDB18-HQ dataset. It contains 40 channels and its segment is 8 seconds. I plan to train it with 240 epoches

CarlGao4 commented 9 months ago

https://drive.google.com/file/d/12t6IhJzOnkRyEBtBx42BaxaUTQYXsCJu/view?usp=drive_link As I didn't train it as much epochs as the official htdemucs models and hdemucs models, it is even worse than the mdx model