created leaky prox (possible remedy for #222)

scarlet/blend.py

@@ -8,6 +8,7 @@
 import logging
 
 from .component import CombinedComponent
+from .constraint import Constraint
 from .model import UpdateException
 
 logger = logging.getLogger("scarlet.blend")
@@ -130,12 +131,20 @@ def expand_grads(*X, func=None):
                     l + p for l, p in zip(grad_logL(*X), grad_logP(*X))
                 )
 
-                # step sizes, allow for random skipping of parameters
+                # step sizes
                 _step = lambda *X, it: tuple(
                     x.step(x, it=it) if hasattr(x.step, "__call__") else x.step
                     for x in X
                 )
-                _prox = tuple(x.constraint for x in X)
+
+                # proxes, allow for random skipping
+                _prox = tuple(
+                    lambda Z, step, parameter=x: 0.95 * parameter.constraint(Z, step)
+                    + 0.05 * Z
+                    if parameter.constraint is not None
+                    else Z
+                    for x in X
+                )
 
                 # good defaults for adaprox
                 scheme = alg_kwargs.pop("scheme", "amsgrad")