alisiahkoohi / langevin-dynamics Goto Github PK

Hi, thanks for the repo! I have three questions regarding the code for Metropolis-adjusted Langevin:

1. When a proposal is rejected, why is the gradient not reset (i.e., self.grad[1].data = self.grad[0].data)?
   

   Langevin-dynamics/langevin_sampling/samplers.py

   Lines 212 to 214 in 495de3a

   	else:
   	self.x[1].data = self.x[0].data
   	self.P[1].data = self.P[0].data

2. When computing the density of the proposal distribution, why isP[1] used for both directions? Shouldn't it be P[idx^1]?
   

   Langevin-dynamics/langevin_sampling/samplers.py

   Lines 228 to 233 in 495de3a

   	return (-(.25 / self.lr_fn(self.counter)) *
   	(self.x[idx] - self.x[idx ^ 1] -
   	self.lr_fn(self.counter) * self.grad[idx ^ 1] / self.P[1]) *
   	self.P[1]
   	@ (self.x[idx] - self.x[idx ^ 1] -
   	self.lr_fn(self.counter) * self.grad[idx ^ 1] / self.P[1]))

3. Why are the rejected samples discarded? Shouldn't MH keep both accepted and rejected samples in the list?
   

   Langevin-dynamics/langevin_sampling/samplers.py

   Line 198 in 495de3a

   while not accepted:

Hi,

Thank you so much for the repository. I am learning MCMC. It's really helpful for me.
I have a question. In class LangevinDynamics(object):, you never run self.optim.zero_grad(). Is this on purpose?

You use the preconditioned SGD in gaussian_LD.py demo. However, if you change pSGLD to SGLD, it actually doesn't work:

Thanks for the code! I have a question regarding the gradient's direction when sampling the next variable.

In class pSGLD(Optimizer), the mean is updated using a negative learning rate p.data.addcdiv_(grad, G, value=-group['lr']). However, when calculating the transition probability in def proposal_dist(self, idx: int) -> torch.Tensor from class MetropolisAdjustedLangevin(object), I think the gradient direction needs to be reversed self.x[idx] - self.x[idx ^ 1] + self.lr_fn(self.counter) * self.grad[idx ^ 1] instead ofself.x[idx] - self.x[idx ^ 1] - self.lr_fn(self.counter) * self.grad[idx ^ 1]. Why is the gradient implemented in another direction?

Hi,

in line 105 of precondSGLD.py, why are you dividing by G to obtain the noise_std? Surely following the paper by Li et al. you should be multiplying by G? Or am I missing something? Because this line is causing my parameter estimates to explode ...
Thanks!