search

keunwoochoi / music-auto_tagging-keras

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

Weights shape for CRNN for TF are incorrect #4

Closed meetps closed 8 years ago

meetps commented 8 years ago

The CRNN takes a 96*1366 spectrogram image as input. However the weights (for input batch normalization layer) provided for Tensorflow for CRNN have the shape of the 

bn_0_freq_beta_1 (1440,)
bn_0_freq_gamma_1 (1440,)
bn_0_freq_running_mean_1 (1440,)
bn_0_freq_running_std_1 (1440,)

which is incompatible with the input image size.

keunwoochoi commented 8 years ago

It is compatible because there is zero padding.