wd 0.002, lr 0.0005, BS 16, nepoch 200
spec = fftweight.fft_gtgram(sig, fs, 1024/fs, 1024/fs, 32, 500)

0.0 : first try, archi 0.0
0.1 : archi 0.0, with white pink noise & rnd shift -> poor results
0.2 : archi 0.2 (only 2 1024 layers)
0.3 : archi 0.3 (only 2 512 layers)

wd 0.002, lr 0.0005, BS 16, nepoch 200
spec = fftweight.fft_gtgram(sig, fs, 1024/fs, 512/fs, 32, 500)
0.3 : archi 0.3, 512 spec hopsize

spec = fftweight.fft_gtgram(sig, fs, 512/fs, 256/fs, 32, 500)
0.4 : archi 0.3, 512 spec winsize

spec = fftweight.fft_gtgram(sig, fs, 512/fs, 256/fs, 32, 2000)
0.5 : archi 0.3, 2000 start freq

spec = fftweight.fft_gtgram(sig, fs, 512/fs, 256/fs, 64, 2000)
0.6 : archi 0.3, 64 freq bins 
0.7 : archi 0.4 (256 features)

spec = fftweight.fft_gtgram(sig, fs, 512/fs, 256/fs, 128, 2000)
0.8 : archi 0.3, 128 freq bins

0.9 same as 0.6, new dataset

0.10 balanced weights



BEST : 0.6 -> 0.88 valid AUC, 0.8 valid ACC


new dataset __
spec = fftweight.fft_gtgram(sig, fs, 512/fs, 128/fs, 64, 2000)
0.61 : archi 0.3 (kernel width = 1) same as 0.6
0.62 : archi 0.5 (kernel width = 11)


0.8 archi 0.6
0.9 archi 0.7, hopsize 128 dataset split = sklearn
0.8 archi 0.8 (kernel sizes 17, 11, 5 insted of 11 11 11), split = BOMBYX2017
0.10 archi 0.7, dataset split test = BOMBYX2017

split BOMBYX2017
0.12 archi 0.7 batchsize 32 -> overfit  plus que 0.11, valid AUC .89
0.11 archi 0.10 batchsize 32  valid AUC .895


LASTS : 
0.13 : gammatone
stft
stftpcen
sincnet