keunwoochoi / music-auto_tagging-keras

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

compact cnn how to obtain the dominant tags ? #32

Closed emmanuelbernieri closed 6 years ago

emmanuelbernieri commented 7 years ago

Hello again,

i've implemented the compact-cnn neural network however the output is a vector of 0 & 1 corresponding to inditifed taggs. Is it possible like in the CRNN to only have the dominant taggs like :

[('jazz', '0.444'), ('instrumental', '0.151'), ('folk', '0.103'), ('Hip-Hop', '0.103'), ('ambient', '0.077')]

Kind regards

Emmanuel

keunwoochoi commented 7 years ago

however the output is a vector of 0 & 1 corresponding to inditifed taggs

Really? I'm using Sigmoid as an activation function in the output layer, so it can't be like that if you're using the result as it is.

emmanuelbernieri commented 7 years ago

Does the music file format has an importance ? I cannot analyse mp3 so I convert them into wav and then I analyse.

after preprocessing the audio input with librosa it gives me an output like that:

[[[ 0. 0. 0. ..., 0.04826343 0.05228163 0.05081845]]]

keunwoochoi commented 7 years ago

Yes, whatever the file format is, it should be decoded which you do when you're using librosa.load.

emmanuelbernieri commented 7 years ago

even when I use the 1100103.clip.npy. The output of the prediction is still [0.....1]

model

emmanuelbernieri commented 7 years ago

I'm using

python 3.6 Keras 2.0.6 Tensorflow 1.3.0 Theano 0.9.0 Librosa 0.5.1

keunwoochoi commented 7 years ago

Please check out the keras version (and possibly others) as readme. I think the theano version was either 0.7 or 0.8 at the moment.

emmanuelbernieri commented 7 years ago

I'll try to set up exactly the same config as yours tomorrow and let you know the result

emmanuelbernieri commented 7 years ago

I didn't succeed in reproducing your result.

Do you have a docker for a demo ?

Kind regards

Emmanuel

keunwoochoi commented 7 years ago

Do you have a docker for a demo ?

No I don’t.

On 7Sep 2017, at 15:18, emmanuelbernieri notifications@github.com wrote:

I didn't succeed in reproducing your result.

Do you have a docker for a demo ?

Kind regards

Emmanuel

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/keunwoochoi/music-auto_tagging-keras/issues/32#issuecomment-327813564, or mute the thread https://github.com/notifications/unsubscribe-auth/APZ8xTcB6CPgq-HjJIViggOKlU1_rSCJks5sf_sdgaJpZM4POIny.

emmanuelbernieri commented 7 years ago

thanks for everything i'll let you know when I find where the problem is