Can't run any of models with the latest python packages

keunwoochoi / music-auto_tagging-keras

Music auto-tagging models and trained weights in keras/theano

MIT License

616 stars 142 forks source link

Can't run any of models with the latest python packages #21

Closed nartes closed 6 years ago

nartes commented 7 years ago

https://gist.github.com/anonymous/062721cae29f62ab7262ceb1d0e66a59 - stack trace https://gist.github.com/anonymous/cfb146880d54675b213e125b1624f313 - requirements.txt https://gist.github.com/anonymous/b30226c691a3d2bd7999b67356dd5461 - keras.conf

I tried both backends - Theano and Tensorflow.

Could you provide your requirements.txt?

I'm using only CPU.

nartes commented 7 years ago

Tensorflow backend can't load model's weights. https://gist.github.com/anonymous/595a4b0670bf0e00410d471ced923166

keunwoochoi commented 7 years ago

Thanks for interest. I am considering removing MusicTaggerCNN/CRNN and replace them with compact_cnn which runs with latest python packages. Could this help you at least for temporary?

nartes commented 7 years ago

Glad to get your reply, Keunwoo!

I just tried to run your implemtation before digging deeper. I need to read your paper, since I don't know what compact_cnn is. Actually I am interested in one of bibliography references from the paper. (https://arxiv.org/abs/1508.01774) I've not found the code for that one, and by chance discovered yours. So I need to learn tensorflow, before that I was playing around with MATLAB only.

keunwoochoi commented 7 years ago

Yeah.. I'd guess cnn/crnn structure for tagging is simpler and more straightforward than those for transcription, just because of the natures of tasks. Numpy/keras/tf is not too hard for ex-MATLAB users!