20 Jan 2017, Keunwoo Choi

• A compact cnn - 5 layers, batch normalizationm, ELU, 32 feature maps for each conv layers.

• AUC is about 0.849 for the tagging task.

• Best of pre-trained: cherry-picked from pre-trained
• Concatenating 12345 + MFCCs : concat(all pre-trained features, MFCCs)
• MFCCs: mean and vars of {MFCC, dMFCC, ddMFCC}
• SoTA: reported state-of-the-art scores

More details coming soon.

• set image_dim_ordering() == th.
• It works on both tensorflow/theano backend.
• install kapre OLD VERSION by

$ git clone https://github.com/keunwoochoi/kapre.git
$ cd kapre
$ git checkout a3bde3e
$ python setup.py install

• See main.py for an example.
• It is not the most efficient implementation, but the easiest for me :) still it's not slow even for cpu-based inference.

Tested on Keras 1.2.1

Coming soon.