cleaned up for submission

unlearning/CertifiedUnlearning.py

@@ -7,7 +7,6 @@ from unlearning.Strategy import Strategy
 from sets.Data import *
 
 # Single-Batch Certified Unlearning for DNNs
-
 class CertifiedUnlearning(Strategy):
     """
     Implements Certified Unlearning for non-convex DNNs (Zhang et al.).
@@ -47,8 +46,7 @@ class CertifiedUnlearning(Strategy):
         params_list = []
         for name, p in inner_model.named_parameters():
             if p.requires_grad:
-                
-                # Append as a tuple so it can be unpacked as (name, param)
+                # so it can be unpacked as (name, param)
                 params_list.append((name, p))
                 
         return params_list if named else [e[1] for e in params_list]
@@ -133,7 +131,7 @@ class CertifiedUnlearning(Strategy):
                         h_estimate[k].copy_(h_estimate[k] + g[k] - (h_s[k] / self.scale))
                         
 
-                # feed back on status
+                # feed back because this took agaes and it was not clear if things were working
                 if global_step % step_interval == 0 and current_pct < 100:
                     current_pct += 1
                     print(f"\rProgress: {current_pct}% done", end="", flush=True)
@@ -167,7 +165,7 @@ class CertifiedUnlearning(Strategy):
                         # Keep early low-level vision filters entirely pristine
                         pass
                 
-                # Move to the next calculated Hessian vector block only after a valid update step
+                # to the next calculated Hessian vector block
                 delta_idx += 1
                 
         return model