search

keunwoochoi / music-auto_tagging-keras

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

What was the GPU that you used to train your model?? #27

Closed kshitijchandna closed 7 years ago

kshitijchandna commented 7 years ago

I am very new to this but which GPU did you use while training it on MTT?? My kernal dies constantly on Tensorflow-GPU with Nvidia 960m.

keunwoochoi commented 7 years ago

I used Titan but it shouldn't be about GPU. Could be about backend - theano and tensorflow. Could you check how's it working if you use compact_cnn folder in the tepo with keras 1.2?

On 14 Apr 2017, at 21:07, kshitijchandna notifications@github.com wrote:

I am very new to this but which GPU did you use while training it on MTT?? My kernal dies constantly on Tensorflow-GPU with Nvidia 960m.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

kshitijchandna commented 7 years ago

You were right it was about the backend. After some minor changes to the code,CRNN worked perfectly with Keras 2 with tensorflow as backend.

keunwoochoi commented 7 years ago

Great. If you're gonna learn your model from scratch, it should be fine.

If you wanna load and use the weights, please note that...

Due to the different BN implementations in keras version <= 1.0.6, and keras <= 1.2.2, and keras 2.x (whoa), the prediction would probably be weird. Compact_CNN is also in the case, so you probably need to use virtualenv and install some old versions. Sorry for this ;)