notes.txt

- ideal goal: be able to say "in this part of the song there are elements of jazz, prog rock and a bit of grunge" :)

- adjust all music samples to be more or less on the same volume level to prevent the model from learning it as a feature

- define distance measure (or similarity measure) between songs / groups

- feed network with more than just mel-spectrogram images?
	-> Q: how much information does mel-spectrogram contain (do faster songs end up not being converted with enough details?)?

- prediction confidence: for arbitrary song (not bounded to 30s), user can demand higher confidence so you create multiple 30s spectrograms


TODO:
- implement measurements as in competition
- plot train/valid graph (30 epochs)
- choose optimal number of epochs for medium dataset
- write submit.py and make successful submission

------------------------------------------------------

- implement 1D convolution
- expand medium dataset