Merge pull request #168 from furukawa-laboratory/#167_fix_linearly_de…

…creasing_sigma_function

#167 近傍半径の線形減少の式を変更する

libs/models/som.py

@@ -66,7 +66,7 @@ def __init__(self, X, latent_dim, resolution, sigma_max, sigma_min, tau, init='r
     def _cooperative_process(self, epoch):
         # 学習量を計算
         # sigma = self.sigma_min + (self.sigma_max - self.sigma_min) * np.exp(-epoch / self.tau) # 近傍半径を設定
-        self.sigma = max(self.sigma_min, self.sigma_max * (1 - (epoch / self.tau)))  # 近傍半径を設定
+        self.sigma = max(self.sigma_min, self.sigma_min + (self.sigma_max - self.sigma_min) * (1 - (epoch / self.tau)))  # 近傍半径を設定
         Dist = dist.cdist(self.Zeta, self.Z, 'sqeuclidean')
         # KxNの距離行列を計算
         # ノードと勝者ノードの全ての組み合わせにおける距離を網羅した行列

libs/models/som_tensorflow.py

@@ -158,7 +158,7 @@ def sigma(self):
 
         with tf.name_scope('Sigma_computation'):
 
-            return tf.maximum(tf.cast(self.sigma_min, tf.float64), self.sigma_max*(1 - (self.iter_no/self.tau)))
+            return tf.maximum(tf.cast(self.sigma_min, tf.float64), self.sigma_min + (self.sigma_max - self.sigma_min) * (1 - (self.iter_no / self.tau)))
 
     def winning_nodes(self):
         """

libs/models/som_use_for.py

@@ -64,7 +64,7 @@ def fit(self, nb_epoch=100, verbose=True):
         # 学習量の計算
             H = np.zeros((self.K,self.N)) # 全ての学習量を保持する配列を用意
             # 近傍半径の計算
-            sigma = max(self.sigma_min, self.sigma_max * ( 1 - (epoch / self.tau) ) )# 線形で減少
+            sigma = max(self.sigma_min, self.sigma_min + (self.sigma_max - self.sigma_min) * (1 - (epoch / self.tau)))# 線形で減少
             #sigma = self.sigma_min + (self.sigma_max - self.sigma_min) * ( 1 - (epoch / self.tau) ) # 指数的に減少
             for k in range(self.K):
                 for n in range(self.N):

libs/models/tsom.py

@@ -145,14 +145,14 @@ def fit(self, nb_epoch=200):
         for epoch in tqdm(np.arange(nb_epoch)):
             # 学習量の決定
             # sigma1 = self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * np.exp(-epoch / self.TAU1)
-            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MAX * (1 - (epoch / self.TAU1)))
+            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * (1 - (epoch / self.TAU1)))
             distance1 = distance.cdist(self.Zeta1, self.Z1, 'sqeuclidean')  # 距離行列をつくるDはN*K行列
             H1 = np.exp(-distance1 / (2 * pow(sigma1, 2)))  # かっこに気を付ける
             G1 = np.sum(H1, axis=1)  # Gは行ごとの和をとったベクトル
             R1 = (H1.T / G1).T  # 行列の計算なので.Tで転置を行う
 
             # sigma2 = self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * np.exp(-epoch / self.TAU2)
-            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MAX * (1 - (epoch / self.TAU2)))
+            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * (1 - (epoch / self.TAU2)))
             distance2 = distance.cdist(self.Zeta2, self.Z2, 'sqeuclidean')  # 距離行列をつくるDはN*K行列
             H2 = np.exp(-distance2 / (2 * pow(sigma2, 2)))  # かっこに気を付ける
             G2 = np.sum(H2, axis=1)  # Gは行ごとの和をとったベクトル

libs/models/tsom2_ishida.py

@@ -144,14 +144,14 @@ def fit(self, nb_epoch=200):
         for epoch in tqdm(np.arange(nb_epoch)):
             # 学習量の決定
             # sigma1 = self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * np.exp(-epoch / self.TAU1)
-            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MAX * (1 - (epoch / self.TAU1)))
+            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * (1 - (epoch / self.TAU1)))
             distance1 = distance.cdist(self.Zeta1, self.Z1, 'sqeuclidean')  # 距離行列をつくるDはK1*N1行列
             H1 = np.exp(-distance1 / (2 * pow(sigma1, 2)))  # かっこに気を付ける
             G1 = np.sum(H1, axis=1)  # Gは行ごとの和をとったベクトル
             R1 = (H1.T / G1).T  # 行列の計算なので.Tで転置を行う
 
             # sigma2 = self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * np.exp(-epoch / self.TAU2)
-            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MAX * (1 - (epoch / self.TAU2)))
+            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * (1 - (epoch / self.TAU2)))
             distance2 = distance.cdist(self.Zeta2, self.Z2, 'sqeuclidean')  # 距離行列をつくるDはK2*N2行列
             H2 = np.exp(-distance2 / (2 * pow(sigma2, 2)))  # かっこに気を付ける
             G2 = np.sum(H2, axis=1)  # Gは行ごとの和をとったベクトル

libs/models/tsom3.py

@@ -130,20 +130,20 @@ def fit(self, nb_epoch=200):
         for epoch in tqdm(np.arange(nb_epoch)):
             # 学習量の決定
             # sigma1 = self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * np.exp(-epoch / self.TAU1)
-            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MAX * (1 - (epoch / self.TAU1)))
+            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * (1 - (epoch / self.TAU1)))
             distance1 = distance.cdist(self.Zeta1, self.Z1, 'sqeuclidean')  # 距離行列をつくるDはK1*N1行列
             H1 = np.exp(-distance1 / (2 * pow(sigma1, 2)))  # かっこに気を付ける
             G1 = np.sum(H1, axis=1)  # Gは行ごとの和をとったベクトル K1*1
             R1 = (H1.T / G1).T  # 行列の計算なので.Tで転置を行う K1*N1
 
             # sigma2 = self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * np.exp(-epoch / self.TAU2)
-            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MAX * (1 - (epoch / self.TAU2)))
+            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * (1 - (epoch / self.TAU2)))
             distance2 = distance.cdist(self.Zeta2, self.Z2, 'sqeuclidean')  # 距離行列をつくるDはK2*N2行列
             H2 = np.exp(-distance2 / (2 * pow(sigma2, 2)))  # かっこに気を付ける
             G2 = np.sum(H2, axis=1)  # Gは行ごとの和をとったベクトル K2*1
             R2 = (H2.T / G2).T  # 行列の計算なので.Tで転置を行う K2*N2
 
-            sigma3 = max(self.SIGMA3_MIN, self.SIGMA3_MAX * (1 - (epoch / self.TAU3)))
+            sigma3 = max(self.SIGMA3_MIN, self.SIGMA3_MIN + (self.SIGMA3_MAX - self.SIGMA3_MIN) * (1 - (epoch / self.TAU3)))
             distance3 = distance.cdist(self.Zeta3, self.Z3, 'sqeuclidean')  # 距離行列をつくるDはK3*N3行列
             # print(distance3.shape)
             H3 = np.exp(-distance3 / (2 * pow(sigma3, 2)))  # かっこに気を付ける

libs/models/tsom_tensorflow.py

@@ -200,7 +200,7 @@ def sigma(self):
         """
 
         with tf.name_scope('Sigma_computation'):
-            return tf.maximum(self.sigma_min * tf.constant(1, tf.float64), self.sigma_max * (1 - (self.iter_no / self.tau)))
+            return tf.maximum(self.sigma_min * tf.constant(1, tf.float64), self.sigma_min + (self.sigma_max - self.sigma_min) * (1 - (self.iter_no / self.tau)))
 
     def winning_nodes(self, M, Y, axis):
         """

libs/models/tsom_use_for.py

@@ -121,7 +121,7 @@ def fit(self, nb_epoch=200):
 
             # mode1の学習量の計算
             #sigma1 = self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * np.exp(-epoch / self.TAU1)
-            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MAX * (1 - (epoch / self.TAU1)))
+            sigma1 = max(self.SIGMA1_MIN, self.SIGMA1_MIN + (self.SIGMA1_MAX - self.SIGMA1_MIN) * (1 - (epoch / self.TAU1)))
             for i in np.arange(self.N1):
                 for k in np.arange(self.K1):
                     zeta_dis1 = 0
@@ -130,7 +130,7 @@ def fit(self, nb_epoch=200):
                     h1[i][k] = np.exp(-0.5 * zeta_dis1 / (sigma1 ** 2))
 
             # mode2の学習量の計算
-            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MAX * (1 - (epoch / self.TAU2)))
+            sigma2 = max(self.SIGMA2_MIN, self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * (1 - (epoch / self.TAU2)))
             #sigma2 = self.SIGMA2_MIN + (self.SIGMA2_MAX - self.SIGMA2_MIN) * np.exp(-epoch / self.TAU2)
             for j in np.arange(self.N2):
                 for l in np.arange(self.K2):

libs/tools/calc_sigma.py

@@ -21,7 +21,7 @@ def __calc_sigma_exp(sigma_max: float, sigma_min: float, tau: float, epoch: int)
 
 
 def __calc_sigma_linear(sigma_max: float, sigma_min: float, tau: float, epoch: int) -> float:
-    return max(sigma_min, sigma_max * (1 - (epoch / tau)))
+    return max(sigma_min, sigma_min + (sigma_max - sigma_min) * (1 - (epoch / tau)))
 
 
 if __name__ == '__main__':