Open xuerq opened 6 years ago
@xuerq
student_predict = np.random.logistic(mu, scale)
p_s = probs_logistic(mu, scale, student_predict)
p_s = Variable(torch.from_numpy(p_s).float()).cuda().permute(0, 2, 1)
# input student sample to evaluate probability
student_predict = Variable(torch.from_numpy(student_predict).float()).cuda()
y_hat = teacher(student_predict) # y_hat: (B x C x T) teacher: 10-mixture-logistic
sorry for replying late,I think z * scale + mu
, z sample from logistic(0,1) and so it seems the same as logistic(mu, scale).
It's a little bit different I think When calculating H(ps(xt | x<t), Pt(xt|x<t)) , “x<t” is fixed。That means,z should be fixed。
If I understand correctly . teacher model's inputs should be z * scale + mu ? https://github.com/zhf459/P_wavenet_vocoder/blob/master/train_student.py#L303