add log=grad as default option to eval

analog/analog.py

@@ -324,8 +324,6 @@ def initialize_from_log(self) -> None:
             lora_state = torch.load(os.path.join(lora_dir, "lora_state_dict.pt"))
             if not is_lora(self.model):
                 self.add_lora(lora_state=lora_state)
-            if not any("analog_lora_A" in name for name in self.model.state_dict()):
-                self.add_lora(lora_state=lora_state)
             for name in lora_state:
                 assert name in self.model.state_dict(), f"{name} not in model!"
             self.model.load_state_dict(lora_state, strict=False)

analog/logging/option.py

@@ -97,11 +97,16 @@ def _sanity_check(self):
             )
             self._log["grad"] = True
 
-    def eval(self):
+    def eval(self, log="grad"):
         """
         Enable the evaluation mode. This will turn of saving and updating
         statistic.
         """
+        if isinstance(log, str):
+            self._log[log] = True
+        else:
+            raise ValueError(f"Unsupported log type for eval: {type(log)}")
+
         self.clear(log=False, save=True, statistic=True)
 
     def clear(self, log=True, save=True, statistic=True):

examples/mnist_influence/compute_influences_scheduler.py

@@ -89,6 +89,6 @@
 
 # Save
 if_scores = if_scores.numpy().tolist()[0]
-torch.save(if_scores, "if_analog_scheduler_init_from_log_0.8.pt")
+torch.save(if_scores, "examples/mnist_influence/if_analog_scheduler.pt")
 print("Computation time:", time.time() - start)
 print("Top influential data indices:", top_influential_data.numpy().tolist())