Reproducing LangevinVAE result in Table 3

[qsh-zh]

Thanks for your insightful paper and sharing code.

I run your shared code on experiment 4 MNIST, I can reproduce results claimed in Table 3 except LangevinVAE. It reported NLL around 200 in multiple runs.

I notice there could be a bug in your code, the line misses y in calling interpolated_force. But the results didn't change even I changed to

x1 = x + stepsize * self.interpolated_force(x, y, lambda_) + torch.sqrt(2*stepsize) * torch.randn_like(x)

[franknoe]

Just spoke with Hao Wu who wrote this code, it's a result of occasional bad proposal step. He will look into details and revise it. Thanks for bringing it up.

[qsh-zh]

Thanks for your timely response. But according to the paper the mean and covariance of LangevinVAE looks not so bad, not sure it is occasional bad proposal step. I try different random seed, and all reports NLL around 200.

Thanks for you again your insightful paper, it is really a nice work.

[amrhamedp]

@franknoe
def energy_1(self, x, y):
return (x**2).sum(axis=2, keepdims=True) / 2 - self.G.conditional_log_likelihood(x, y).sum(axis=2, keepdims=True)
The bug is because lambda is passed instead of y and conditional log likelihood is calculated using the lambda value.

[franknoe]

thank you! Could you send a PR?

[carsonlawrence]

Hi there, I’m just following the bread crumbs on this and I’m a bit confused as to the status of the Langevin VAE result in Table 3. Is the result in the paper correct? How should I reproduce it, please? Thanks