What is the difference between sample() and rsample()?

Question

When I sample from a distribution in PyTorch, both sample and rsample appear to give similar results:

import torch, seaborn as sns

x = torch.distributions.Normal(torch.tensor([0.0]), torch.tensor([1.0]))

sns.distplot(x.sample((100000,)))	sns.distplot(x.rsample((100000,)))

When should I use sample(), and when should I use rsample()?

Answer 1

Using rsample allows for pathwise derivatives:

The other way to implement these stochastic/policy gradients would be to use the reparameterization trick from the rsample() method, where the parameterized random variable can be constructed via a parameterized deterministic function of a parameter-free random variable. The reparameterized sample therefore becomes differentiable.