About equation (9) in the paper.

[YukunLi99]

I find that the score is taken as the opposite number in the code. Why is the final score calculated in this way? The equation (9) in the paper is: The following is the code implementation:

composition = -torch.logsumexp(torch.stack((-e1, -e2), dim=0), dim=0)

[linhaowei1]

Hello,

This might be a legacy typo. The composition score simulating the OR operator uses the function $-logsumexp(-e_1-e_2)$ in [1], while we adopt a different definition of energy functions $p(x) = e^{E(x)}$ to align with the definition of the OOD score.

In our experiments, I recall that the results using torch.logsumexp(torch.stack((e1, e2), dim=0), dim=0) and -torch.logsumexp(torch.stack((-e1, -e2), dim=0), dim=0) were similar (though I am not completely certain as it has been some time since then). If you are interested, feel free to give it a try.

[1] Du et al. Compositional Visual Generation with Energy Based Models. NeurIPS 2020.