回答編集履歴

fix cnn

2023/01/26 07:17

投稿

スコア1579

test CHANGED Viewed

@@ -205,27 +205,21 @@
 ]
 def cnn(x):
+    x = Conv2D(64, 5, **kwargs)(x)
+    x = BatchNormalization()(x)
+    x = Activation("swish")(x)
+    x = DepthwiseConv2D(4, 2, **kwargs)(x)
+    x = BatchNormalization()(x)
+    x = Activation("swish")(x)
-    for layer in cnn_layers:
+    for layer in cnn_layers:
-        x = layer(x)
+        x = layer(x)
-    return x
+    return x
-inputA = Input(shape = (512, 512, 1), name = "input_a")
+inputA = Input(shape = (512, 512, 1), name = "input_a")
-x = Conv2D(64, 5, **kwargs)(inputA)
-x = BatchNormalization()(x)
-x = Activation("swish")(x)
-x = DepthwiseConv2D(4, 2, **kwargs)(x)
-x = BatchNormalization()(x)
-x = Activation("swish")(x)
-x = cnn(x)
+x = cnn(inputA)
-inputB = Input(shape = (512, 512, 1), name = "input_b")
+inputB = Input(shape = (512, 512, 1), name = "input_b")
-y = Conv2D(64, 5, **kwargs)(inputB)
-y = BatchNormalization()(y)
-y = Activation("swish")(y)
-y = DepthwiseConv2D(4, 2, **kwargs)(y)
-y = BatchNormalization()(y)
-y = Activation("swish")(y)
-y = cnn(y)
+y = cnn(inputB)
 z = Concatenate(axis = -1)([x, y])
 z = Dropout(0.25)(z)

fix answer

2023/01/26 07:05

投稿

スコア1579

test CHANGED Viewed

@@ -150,7 +150,7 @@
 これでパラメータ数が`55,338,485`まで更に54%削減できました．当初のモデルと比較して95%もの削減になっています．
 ## 削減案4
-流行を取り入れて次のようにします．`Dense`が5段あるのは多い気がしますが`Dropout`で過学習をカバーします．
+`inputA`と`inputB`に大きな特徴の差異があっても良いように，序盤の畳み込みは各個別で行った上で流行を取り入れて次のようにします．`Dense`が5段あるのは多い気がしますが`Dropout`で過学習をカバーします．
 ```Python
 from tensorflow.keras.layers import Input, Dense, Flatten, Concatenate, Activation

fix dense initializer

2023/01/26 07:01

投稿

スコア1579

test CHANGED Viewed

@@ -201,7 +201,7 @@
     BatchNormalization(),
     Activation("swish"),
     GlobalAveragePooling2D(),
-    Dense(1024, activation = "swish"),
+    Dense(1024, activation = "swish", kernel_initializer = "he_uniform"),
 ]
 def cnn(x):
@@ -229,11 +229,11 @@
 z = Concatenate(axis = -1)([x, y])
 z = Dropout(0.25)(z)
-z = Dense(2048, activation = "swish")(z)
+z = Dense(2048, activation = "swish", kernel_initializer = "he_uniform")(z)
 z = Dropout(0.25)(z)
-z = Dense(1024, activation = "swish")(z)
+z = Dense(1024, activation = "swish", kernel_initializer = "he_uniform")(z)
 z = Dropout(0.25)(z)
-z = Dense(256, activation = "swish")(z)
+z = Dense(256, activation = "swish", kernel_initializer = "he_uniform")(z)
 output = Dense(class_num, activation = "softmax", name = "output")(z)
 model = Model([inputA, inputB], output)