Dimensionality change

[jkamalu]

In the tacotron paper, the expected dimensionality of the tensor during most of the encoder cbhg module is expected to be batch size x time steps x num features where num features = 128. Why do you shift to 256 before the bidirectional GRU layer? It seems like this would result in a real loss of information during encoding. Is this somehow what is described in the paper and I am just missing it? Thanks

[Spotlight0xff]

Yes, I noticed that as well.
I believe I fixed that with my PR.
Look again at the code, seems fine to me now.