CausalVAE: no update of lambda and c in the Lagrangian loss approximation of the DAGness?

[azlaanmsamad]

The paper mentions in the appendix equation 19:
we define an augmented Lagrangian:
l_pre = l_u + λ*H(A) + (c^2/ 2) * H^2(A)

and the update equation for c and λ. But you could not find this in the code, instead c and λ are hardcoded requiring no update?

L, kl, rec, reconstructed_image,_ = lvae.negative_elbo_bound(u,l,sample = False)
dag_param = lvae.dag.A
h_a = _h_A(dag_param, dag_param.size()[0])
L = L + 3*h_a + 0.5*h_a*h_a #- torch.norm(dag_param)
L.backward()

Is there are a reason for this or it has been mistakenly left out?

[shaido987]

Hello,

The first author of CausalVAE has already completed her internship with us. For a faster and more detailed reply, you can email this and any further questions to Mengyue Yang. You could include a link to this issue as context.