Hello, I think some of the unreasonable about your test set settings

andabi / music-source-separation

Deep neural networks for separating singing voice from music written in TensorFlow

796 stars 149 forks source link

Hello, I think some of the unreasonable about your test set settings #12

Open 22w2w2 opened 6 years ago

22w2w2 commented 6 years ago

Your project is great, but I still have some confusion. In the testing phase, you still use _sample_range when reading the file, and np.random.choice in _sample_range. Is this not a bad idea? Because in the test we should test all the speech sequences, rather than randomly taking a starting position, very much looking forward to your answer

ShengleiH commented 6 years ago

@22w2w2 I agree with you, and I am confused about the bss_eval metric used here, there are music and vocal "gndsr, gsir, gsar", but I think only the vocal part is reasonable.