23-10-21

loss + mse 로 조기 수렴 시 초기화 적용
파티클의 초기화를 opeimizer 에서 particle 객체로 변경
메모리의 점진적인 누수 #6 현재 누수가 다시 조금씩 증가하는것이 보임

pso/particle.py

@@ -38,7 +38,7 @@ class Particle:
         self.loss = loss
 
         try:
-            if converge_reset and converge_reset_monitor not in ["acc", "accuracy", "loss"]:
+            if converge_reset and converge_reset_monitor not in ["acc", "accuracy", "loss", "mse"]:
                 raise ValueError(
                     "converge_reset_monitor must be 'acc' or 'accuracy' or 'loss'"
                 )
@@ -50,10 +50,12 @@ class Particle:
             print(e)
             exit(1)
 
-        self.reset_particle()
+        self.__reset_particle__()
+        self.best_weights = self.model.get_weights()
+        self.before_best = self.model.get_weights()
         self.negative = negative
         self.mutation = mutation
-        self.best_score = 0
+        self.best_score = [np.inf, 0, np.inf]
         self.before_w = 0
         self.score_history = []
         self.converge_reset = converge_reset
@@ -131,14 +133,20 @@ class Particle:
             (float): 점수
         """
         score = self.model.evaluate(x, y, verbose=0)
-        if renewal == "acc":
-            if score[1] > self.best_score:
-                self.best_score = score[1]
+        if renewal == "loss":
+            if score[0] < self.best_score[0]:
+                self.best_score[0] = score[0]
                 self.best_weights = self.model.get_weights()
-        elif renewal == "loss":
-            if score[0] < self.best_score:
-                self.best_score = score[0]
+        elif renewal == "acc":
+            if score[1] > self.best_score[1]:
+                self.best_score[1] = score[1]
                 self.best_weights = self.model.get_weights()
+        elif renewal == "mse":
+            if score[2] < self.best_score[2]:
+                self.best_score[2] = score[2]
+                self.best_weights = self.model.get_weights()
+        else:
+            raise ValueError("renewal must be 'acc' or 'loss' or 'mse'")
 
         return score
 
@@ -156,6 +164,11 @@ class Particle:
             self.score_history.append(score[1])
         elif monitor in ["loss"]:
             self.score_history.append(score[0])
+        elif monitor in ["mse"]:
+            self.score_history.append(score[2])
+        else:
+            raise ValueError(
+                "monitor must be 'acc' or 'accuracy' or 'loss' or 'mse'")
 
         if len(self.score_history) > patience:
             last_scores = self.score_history[-patience:]
@@ -163,19 +176,18 @@ class Particle:
                 return True
         return False
 
-    def reset_particle(self):
+    def __reset_particle__(self):
         self.model = keras.models.model_from_json(self.model.to_json())
-        self.model.compile(optimizer="adam", loss=self.loss,
-                           metrics=["accuracy"])
-        init_weights = self.model.get_weights()
-        i_w_, i_s, i_l = self._encode(init_weights)
+        self.model.compile(
+            optimizer="adam",
+            loss=self.loss,
+            metrics=["accuracy", "mse"]
+        )
+        i_w_, i_s, i_l = self._encode(self.model.get_weights())
         i_w_ = np.random.uniform(-0.05, 0.05, len(i_w_))
         self.velocities = self._decode(i_w_, i_s, i_l)
 
-        self.best_weights = init_weights
-        self.before_best = init_weights
-
-        del init_weights, i_w_, i_s, i_l
+        del i_w_, i_s, i_l
         self.score_history = []
 
     def _update_velocity(self, local_rate, global_rate, w, g_best):
@@ -212,7 +224,7 @@ class Particle:
                 + -1 * global_rate * r_1 * (encode_g - encode_w)
             )
             if len(self.score_history) > 10 and max(self.score_history[-10:]) - min(self.score_history[-10:]) < 0.01:
-                self.reset_particle()
+                self.__reset_particle__()
 
         else:
             new_v = (
@@ -324,7 +336,7 @@ class Particle:
 
         if self.converge_reset and self.__check_converge_reset__(
                 score, self.converge_reset_monitor, self.converge_reset_patience, self.converge_reset_min_delta):
-            self.reset_particle()
+            self.__reset_particle__()
 
         return score