Training not converged

[JesseYang]

I train the network with more steps, and found that the loss increases around about 9k steps, as follows:

Since the VCTK-Corpus has about 44k wave files, and the batch size is 1, the loss increases during the first epoch.

[ibab]

Not sure why the loss starts increasing in this case, maybe it had something to do with the bug in #29 .
Note that the default hyperparameters are pretty much randomly chosen.
It might be useful to enable batching and use a different learning rate.