czm0/draw_pytorch

DRAW: A Recurrent Neural Network For Image Generation

draw_pytorch's Issues

Bug in filterbank matrices

Hi, you have bug in filterbank matrices computation.

https://github.com/chenzhaomin123/draw_pytorch/blob/master/draw_model.py

        temp,t_sigma = align(temp,sigma2)
        temp = temp / (t_sigma * 2)
        F = torch.exp(-torch.pow(temp,2))

Here you are squaring the sigma2, which is already squared.
Additionally you are squaring the 2, which is less of a problem, but still :)

loss function in draw_model.py

Lz - self.T / 2 should not be contained in the loop, i.e. the loss function should be calculated use your commented code like below:
for t in range(self.T):
mu_2 = self.mus[t] * self.mus[t]
sigma_2 = self.sigmas[t] * self.sigmas[t]
logsigma = self.logsigmas[t]
Lz += (0.5 * (mu_2 + sigma_2 - 2 * logsigma))
Lz -= self.T / 2