Sampling with likelihoods #79
Conversation
mhavasi
left a comment
mhavasi
left a comment
There was a problem hiding this comment.
Hey, thanks for the PR. Can you also add a test that compares the output of the forward compute_likelihood and the backward compute_likelihood to make sure they match?
timonpalm
force-pushed
the
sample_w_likelihood
branch
from
1172ef7 to
0e04528
Compare
| torch.allclose(forward_log_likelihood, backward_log_likelihood, atol=1e-2), | ||
| ) | ||
|
|
||
| def test_forward_backward_likelihoods(self): |
There was a problem hiding this comment.
Hey, thanks for the PR. Can you also add a test that compares the output of the forward compute_likelihood and the backward compute_likelihood to make sure they match?
Is this sufficient? It computes the forward likelihood, while generating samples. Then using those samples to compute the backward likelihood and compare them.
The odesolver.compute_likelihood does not allow a forward computation, as it also would't make sense without x1 samples?!
Maybe I am misunderstanding sth. here ...
|
Hey @timonpalm , thanks for the reply. Can you rebase your branch to main so we can run the unit tests? We just added the option to run unit tests on external PRs recently. |
timonpalm
force-pushed
the
sample_w_likelihood
branch
from
1323874 to
5d778c9
Compare
timonpalm
force-pushed
the
sample_w_likelihood
branch
from
5d778c9 to
4404acb
Compare
2 participants
Issue #62
I expanded the
ODESolver.sample()method to compute likelihoods along the way when sampling, alleviating the two step integration.You just have to pass the
log_p0parameter to the function.