conditioning on z, additive residual connection on each decoder layer from the previous layer

concat 6 dim one-hot on z z_dim = 36

raw wave -> downsample -> min max normalization over 6 hex pickups -> CQT -> log scale -> z score normalize

wave -> cqt -> db cqt -> z score normalization might improve after more epochs changing the lr would help

cqt -> abs -> db -> z score normalize

tend to overfit after 50 epochs

results/20220922_002034_baseline

hop 512 epoch 192 no normalization abs(CQT) as input

attention dropout = 0 input feature = abs(cqt) # no normalization or scaling 64 -> 128 epoch lr = 0.0001