Closed calverca closed 5 years ago
Nope, that's normal. You have to reach 0,1's loss at least to have a full music track
Thank you. I will do that. Appreciate the feedback.
It works. I have also lowered all the LSTM to 256 and increased the batches to 164 and it works nice now. I also found that increasing the difference between the training files helped a lot. I used a metal song and a jazz song for training and got much faster results. Thanks again!
Hello, I have run the files and my test_output file is just one note repeating. I ran 20 epochs with 3 files the loss decreased from 4.3 to 3.9. In the predict.py file I manually replaced the weights.hdf5 with the last weight-improvement hdf5 file. Is there something I'm doing wrong?