Question about calculate_kl

[longcwang3]

For calculate_kl, I am assuming mu_q refers to the variational posterior and mu_p refers to the prior.

PyTorch-BayesianCNN/metrics.py

Lines 27 to 29 in d93bad5

	def calculate_kl(mu_q, sig_q, mu_p, sig_p):
	kl = 0.5 * (2 * torch.log(sig_p / sig_q) - 1 + (sig_q / sig_p).pow(2) + ((mu_p - mu_q) / sig_p).pow(2)).sum()
	return kl

While in the following codes, prior_mu is input as mu_q.

PyTorch-BayesianCNN/layers/BBB_LRT/BBBLinear.py

Lines 75 to 79 in d93bad5

	def kl_loss(self):
	kl = KL_DIV(self.prior_mu, self.prior_sigma, self.W_mu, self.W_sigma)
	if self.use_bias:
	kl += KL_DIV(self.prior_mu, self.prior_sigma, self.bias_mu, self.bias_sigma)
	return kl

Could you check this?