in the real NVP paper they state that the log determinant of the jacobian e ^ sum(of diag entries) which in itself is just the product of those values exponentiated. However in your implementation you do neither.
You don't take the exponent nor do you compute the product you only do the following
log_scale, shift = self.net(x_masked).chunk(2, dim=1)
log_scale = log_scale.tanh() * self.scale_scale + self.shift_scale
log_scale = log_scale * (1 - self.mask)
...
return log_scale
am I missing something here ?
Thank you for the nice videos
in the real NVP paper they state that the log determinant of the jacobian e ^ sum(of diag entries) which in itself is just the product of those values exponentiated. However in your implementation you do neither.
You don't take the exponent nor do you compute the product you only do the following
am I missing something here ?
Thank you for the nice videos