Skip to content

Puzzles about Inconsistency between code and article #3

Description

@QJ-Chen
# losses.sliced_sm
def sliced_score_estimation(score_net, samples, n_particles=1):
    dup_samples = samples.unsqueeze(0).expand(n_particles, *samples.shape).contiguous().view(-1, *samples.shape[1:])
    dup_samples.requires_grad_(True)
    vectors = torch.randn_like(dup_samples)
    vectors = vectors / torch.norm(vectors, dim=-1, keepdim=True)

    grad1 = score_net(dup_samples)  # H, estimation of score
    gradv = torch.sum(grad1 * vectors)  # project H with v
    loss1 = torch.sum(grad1 * vectors, dim=-1) ** 2 * 0.5  # second term of J(\theta) 
    grad2 = autograd.grad(gradv, dup_samples, create_graph=True)[0] # grad of h w.r.t samples(z)
    loss2 = torch.sum(vectors * grad2, dim=-1)

    loss1 = loss1.view(n_particles, -1).mean(dim=0)
    loss2 = loss2.view(n_particles, -1).mean(dim=0)

    loss = loss1 + loss2
    return loss.mean(), loss1.mean(), loss2.mean()

# losses.vae.elbo_ssm
z = imp_encoder(X)
ssm_loss, *_ = sliced_score_estimation_vr(functools.partial(score, dup_X), z, n_particles=n_particles)

To my understanding, grad1 is the estimation of score $h = S_{m}(x;\theta)$ and loss2 is the first term of $J(\theta)$, which is $v^{T}\nabla_{x}h(x;\theta)v$. But in the code, it seems to be calculated as $v^{T}\nabla_{z}h(x;\theta)v$.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions