search

BShakhovsky / PolyphonicPianoTranscription

Recurrent Neural Network for generating piano MIDI-files from audio (MP3, WAV, etc.)
https://magenta.tensorflow.org/onsets-frames
230 stars 41 forks source link

Accuracy stays at zero #7

Closed nikisarv closed 2 years ago

nikisarv commented 3 years ago

Hello,

I'm trying to run your code but during training accuracy dixon does not increase. 936/2212 [===========>..................] - ETA: 53s - loss: 0.0374 - Dixon: 0.0000e+00 944/2212 [===========>..................] - ETA: 52s - loss: 0.0373 - Dixon: 0.0000e+00

I also had a problem with categorical_crossentropy and after i saw your response in another issue i changed it to binary_crossentropy but now i get the results above. I tried it both with the whole dataset and just a small part of it and nothing changes.

BShakhovsky commented 3 years ago

Hi,

That is strange. I just tried running the template, and "Dixon" accuracy is decreasing at first, but starts increasing very soon. I waited for around 1000 steps as in your example, and loss is also around 0.037, but "Dixon" is already above 0.16 by that time. No idea comes to my mind, what may be the reason in your case. Sorry.

nikisarv commented 3 years ago

Hello again and thank you for your reply!

The minSecs and maxSecs you are using for your results is the 5 and 20 right?

Also, I'm using an older version of your code (with tf-gpu 1.15) because when I've tried the newer version I had a memory leak during model.fit(). I'm trying to fix this now and run the newer one. Could you please tell me which keras version you are using along with tensorflow 2.4 & python 3.8.8?

Thank you in advance, NIki

BShakhovsky commented 3 years ago

Yes, currently minSecs and maxSecs are 5 and 20. But I don't think, different values could cause zero accuracy. My Keras version is 2.4.3.

nikisarv commented 3 years ago

Thanks again for your quick reply

I will keep looking for the problem resulting in zero accuracy in my case.