fix merge

src/mrpro/algorithms/optimizers/__init__.py

@@ -4,4 +4,4 @@
 from mrpro.algorithms.optimizers.lbfgs import lbfgs
 from mrpro.algorithms.optimizers.pgd import pgd
 
-__all__ = ["OptimizerStatus", "adam", "cg", "lbfgs", "pdg"]
+__all__ = ["OptimizerStatus", "adam", "cg", "lbfgs", "pdg", "pgd"]

src/mrpro/algorithms/optimizers/pgd.py

@@ -4,28 +4,12 @@
 
 import torch
 
-from mrpro.operators.Functional import Functional
-
-
-def grad_and_value(function, x, create_graph=False):
-    def inner(x):
-        if create_graph and x.requires_grad:
-            # shallow clone
-            xg = x.view_as(x)
-
-        xg = x.detach().requires_grad_(True)
-        (y,) = function(xg)
-
-        yg = y if isinstance(y, torch.Tensor) else y[0]
-        grad = torch.autograd.grad(yg, xg)[0]
-        return grad, y
-
-    return inner(x)
+from mrpro.operators.Functional import Functional, ProximableFunctional
 
 
 def pgd(
     f: Functional,
-    g: Functional,
+    g: ProximableFunctional,  # TODO: would it work with ProximableFunctionalSeparableSum?
     initial_value: torch.Tensor,
     stepsize: float = 1.0,
     reg_parameter: float = 0.01,
@@ -97,7 +81,7 @@ def pgd(
     for _ in range(max_iterations):
         while stepsize > 1e-30:
             # calculate the proximal gradient step
-            gradient, f_y = grad_and_value(f, y)
+            gradient, f_y = torch.func.grad_and_value(f, y)
             (x,) = g.prox(y - stepsize * gradient, reg_parameter * stepsize)
 
             if not backtracking:

tests/algorithms/test_pgd.py

@@ -1,10 +1,10 @@
 """Tests for the proximal gradient descent."""
+
 import torch
 from mrpro.algorithms.optimizers import pgd
 from mrpro.data.SpatialDimension import SpatialDimension
 from mrpro.operators import FastFourierOp
-from mrpro.operators.functionals import L1Norm
-from mrpro.operators.functionals import L2NormSquared
+from mrpro.operators.functionals import L1Norm, L2NormSquared
 from mrpro.phantoms import EllipsePhantom