keunwoochoi / music-auto_tagging-keras

Music auto-tagging models and trained weights in keras/theano
MIT License
616 stars 142 forks source link

parameter problem #7

Closed gongqiang closed 7 years ago

gongqiang commented 7 years ago

the input is 96X1366, and as we multiply all the pooling param in that dimension, we get: 96=2x2x2x3x4, but in another dimension it doesn't fit at all: 4x4x4x5x4=1280,shouldn't it be1366?

keunwoochoi commented 7 years ago

You're correct. A better way would be either used shorter clips from the beginning - or pad zeros to use all 1366 frames. The current setting is chosen just because it was convenient to do so.

gongqiang commented 7 years ago

Thanks~