So I got wavenet working.
Well, for a limited definition of “working.” I’ve been running some experiments to try to figure out how the model behaves and unfortunately I don’t seem to have a very good grasp of the underlying theory of the network architecture, because my experiments have gotten progressively worse.
- The first sample was using the default settings for the implementation I’m using.
- The second sample was the same except for longer training.
- The third and fourth were the same as the previous except with a larger network (more layers).
- For the fifth I changed the activation function and the optimizer to something I thought might work better. (It didn’t.)
- And the sixth I tried training with l2 normalization on.
To be honest, these results are somewhat disheartening, given that I haven’t really had much success, but I’m really interested in the idea so I’m definitely going to stay at it, because I’ve seen some really amazing results from other people experimenting with it.