Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Use Hessian formulation of Fisher information in make_empirical_fisher_vp #430

Merged
merged 4 commits into from
Dec 8, 2023
Merged

Use Hessian formulation of Fisher information in make_empirical_fisher_vp #430

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
a4dccb9
hessian vector product formulation for fisher
agrawalraj Dec 7, 2023
02bf088
ignoring small type error
agrawalraj Dec 7, 2023
82e2768
fixed linting error
agrawalraj Dec 7, 2023
5bb4ce7
Merge branch 'staging-robust' of github.com:BasisResearch/causal_pyro…
agrawalraj Dec 8, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 11 additions & 7 deletions chirho/robust/internals/linearize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,17 +94,21 @@ def make_empirical_fisher_vp(
randomness="different",
)

N = data[next(iter(data))].shape[0] # type: ignore
mean_vector = 1 / N * torch.ones(N)

def bound_batched_func_log_prob(params: ParamDict) -> torch.Tensor:
return batched_func_log_prob(params, data)

def jvp_fn(v: ParamDict) -> torch.Tensor:
return torch.func.jvp(bound_batched_func_log_prob, (log_prob_params,), (v,))[1]

vjp_fn = torch.func.vjp(bound_batched_func_log_prob, log_prob_params)[1]

def _empirical_fisher_vp(v: ParamDict) -> ParamDict:
jvp_log_prob_v = jvp_fn(v)
return vjp_fn(jvp_log_prob_v / jvp_log_prob_v.shape[0])[0]
def jvp_fn(log_prob_params: ParamDict) -> torch.Tensor:
return torch.func.jvp(
bound_batched_func_log_prob, (log_prob_params,), (v,)
)[1]

# Perlmutter's trick
vjp_fn = torch.func.vjp(jvp_fn, log_prob_params)[1]
return vjp_fn(-1 * mean_vector)[0] # Fisher = -E[Hessian]

return _empirical_fisher_vp

Expand Down
Loading