First Fragment - Gloire Éternelle (AI Instrumental, work in progress)
28Collected a larger dataset and have been training a new vocal remover. Also came up with a new implementation of convolutional multihead attention that seems to work a lot better (need to do more thorough tests). This version is only at the 5th epoch, so I'm quite hopeful that it will converge on nearly perfect quality by the end of training.
This isn't final but is of sufficient quality that I felt ok posting it. If you listen closely you can hear vocals at times. Currently working on yet another dataset augmentation that uses a similar neural network to extract vocal stems from instrumental + mix song combinations which will then allow me to create even more training data on the fly. Currently at 2.4 billion training item combinations so perhaps this is a bit unnecessary but it should be fun nonetheless (and can let me extract vocal stems from First Fragment songs!)
The convolutional multihead attention is now using convolutions which share a kernel for each channel for the query, key, value, and out projections so it is more in line with the original multihead attention with its shared linear projections. To do this I just coded a basic shared kernel convolution module; it has weights of size 1x1xHxW and then its expanded in the forward pass to have size Cx1xHxW and then uses a separable convolution. Eventually will get around to coding a pytorch extension for this (unless there already is one I'm unaware of)
If anyone in First Fragment wants me to take this down just let me know, you'll get monetization from this but I understand if you don't want people posting this kind of thing. Goes without saying but I love this song both with and without the vocals, so nothing against them whatsoever.