Closed Merlin-721 closed 3 years ago
I am training with a batch size of 64 (hence seemingly low step count) and LRs of 0.0005.
This looks like overfitting. You could check the training data size, the validation code, etc.
Line 117 in data_loader.py:
a = np.clip(a, 0, 1)
sets all the spectrograms to 0, since my preprocessing uses mostly negative values... that took so long to find!
I am attempting to retrain at 22050Hz. At this SR validation loss for G and P do not decrease (P actually steadily increases). I am using test samples from every speaker in train set. Both loss_id's decrease as expected.
I train G according to this code in solver.py:
and train P according to this code:
I feel this may be due to the LSTMs in the encoders and decoders, since at a different SR the vocal features appear over a different scale, however any other suggestions would be appreciated.