Hi, I have been a problem when trying to train the network with spectrograms creates from audio streaming in real-time. I collected 8k images, 4k for each folder. The spectrograms didn't create with .wav files. It is the only different about your approach.
Do you have any idea about this? My spectrogram images have a resolution of 553x396.
im running data maker to make spectrograms out of ~3000 wav files. for some reason after the spectrogram is made and saved to the file system its RAM memory isn't released. I need to restart the program every ~400 conversions because its becoming incredibly slow.