What's the difference of b6 and M?

[MonolithFoundation]

does they in same performance?

[vanIvan]

The difference is in naming + configuration, M model is of same size / flops as b3-b4 models. After initial experiments, we decided to stick to lesser number of configurations: S (b2), M (~b3-b4), L (b5-b6). The main hyperparameter that plays role in models size and complexity is C - initial number of channels, it is set to:

• 16 for S model
• 24 for M model
• 32 for L model

Also for new model configurations we set hop length of melbanks to 160, decreasing from 200 - it made models a bit heavier in terms of FLOPs

[MonolithFoundation]

Hi, i noticed that there are various sizes available, but for largest one, namely b6 or maybe L, doesn't have option trained on all dataset same as M, would consider add support for this? The most large one and most dataset trained. Some scenarios needs accuracy over speed, also want to see how's the bar could be.

[bzikst]

In my scenarios M model performs better in some cases in comparison to b6. It would be nice to see L model. And, btw, thank you for a great work, I love this models. Waiting for the next steps in your work.

[MonolithFoundation]

@IDRnD could u consider this?

[vanIvan]

I'm sorry, but we currently have suspended all work on redimnet models, due to lack of computational resources.