Archspire - Bleed the Future - 04 Bleed the Future (Instrumental WIP)
17Trained a hierarchical version of my vocal remover, this version is a bit old and I currently have a new version training, but this seems to better in some areas than the previous videos I've posted. Obviously some vocals bleed through but I haven't really committed to training any of these very far, these are all at like the 12th epoch at most and need to be trained a lot longer. Currently have a new version training that will likely be a lot better, will post final versions of all of these videos eventually. Had someone request this, probably won't be posting more Archspire unless people want me to.
You can think of these videos as teasers, I will be uploading final versions of all of these at some point with a model that is trained more fully.
This version is using a VoVNet2 architecture (which changed back to a DenseNet in my current version) where each block corresponds to a specific resolution spectrogram, obviously took some inspiration from Nvidia's hierarchical GANs. It begins with just stage 1 which deals with low resolution spectrograms then builds up new stages for dealing with higher resolution spectrograms while also deepening the encoder/decoder sequences for previous stages in order for them to operate on the same resolution. All modules from previous resolutions are frozen and are not optimized further - interestingly enough I've found it doesn't usually help to optimize those during later stages and can actually hurt.
Currently have a new version training which doesn't use any attention mechanism, I suspect that is part of the reason for so much dulling. It was a fun idea but I don't think it works super well for this task...