[FIX] Examples

docs/examples/basic_usage/example_fisher_monte_carlo.py

@@ -79,7 +79,7 @@
 # identity matrix:
 
 GGN = GGNLinearOperator(model, loss_function, params, data).to_scipy()
-F = FisherMCLinearOperator(model, loss_function, params, data)
+F = FisherMCLinearOperator(model, loss_function, params, data).to_scipy()
 
 D = GGN.shape[0]
 identity = numpy.eye(D)
@@ -109,7 +109,9 @@
 residual_norms = []
 
 for mc in mc_samples:
- F = FisherMCLinearOperator(model, loss_function, params, data, mc_samples=mc)
+ F = FisherMCLinearOperator(
+ model, loss_function, params, data, mc_samples=mc
+ ).to_scipy()
  F_mat = F @ identity
  residual_norms.append(numpy.linalg.norm(GGN_mat - F_mat))
 
@@ -131,12 +133,16 @@
 # first linear operator, we generate some random numbers to show that the
 # global random number generator does not influence the Monte-Carlo estimator:
 
-F1_mat = FisherMCLinearOperator(model, loss_function, params, data) @ identity
+F1_mat = (
+ FisherMCLinearOperator(model, loss_function, params, data).to_scipy() @ identity
+)
 
 # draw some random numbers to modify the global random number generator's state
 torch.rand(123)
 
-F2_mat = FisherMCLinearOperator(model, loss_function, params, data) @ identity
+F2_mat = (
+ FisherMCLinearOperator(model, loss_function, params, data).to_scipy() @ identity
+)
 
 # still, we get the same deterministic approximation
 residual_norm = numpy.linalg.norm(F1_mat - F2_mat)
@@ -201,7 +207,7 @@
  # linear operators indeed realize deterministic matrices.
  F = FisherMCLinearOperator(
  model, loss_function, params, data, seed=seed, check_deterministic=False
- )
+ ).to_scipy()
  F_accumulated += F @ identity
  if mc + 1 in mc_samples:
  F_snapshots.append(F_accumulated / (mc + 1))