Strange query input shape of multihead attention in STL

[elch10]

When I try melltoron with different hparams found strange line in STL module: https://github.com/NVIDIA/mellotron/blob/master/modules.py#L94

d_q = hp.token_embedding_size // 2

So, in hparams file I found that ref_enc_gru_size is half of token_embedding_size, so all work normally without changing hparams: https://github.com/NVIDIA/mellotron/blob/master/hparams.py#L97 But If you try to change ref_enc_gru_size to another value script will raise error. Right shape is:

d_q = hp.ref_enc_gru_size

[rafaelvalle]

Good catch, thank you!