Use COEF value.

[p0p4k]

reference : https://github.com/facebookarchive/loop/blob/112975599b1838a33f139a6c6df0fcd9953ee33f/model.py#L117
alpha_t = self.COEF * torch.sum(phi_t, 1)

[p0p4k]

Ok, maybe i am slightly confused now.
https://arxiv.org/pdf/1906.01083.pdf [MELNET] <-- graves attention is implemented according to this paper in our repo as stated in this discussion where tacotron2 was first released in mozillatts. (mozilla/TTS#346).
However, this paper [VoiceLoop] (https://arxiv.org/pdf/1707.06588.pdf) , Page 5, equation 3 has slightly different equation.
I am trying to match the sigmoid with exp, but it doesn't work. The sigma under square root is missing in our implementation, and so is the 1/sqrt(2*pi) [which is the COEF].