[CNN] Global Average Pooling 層のすすめ

Global Average Pooling 層の良いポイント

パラメーター数を非常に少なくすることができる

→　モデルが単純になり、過学習をしにくくなる

Flatten 層と Global Average Pooling 層の比較

Flatten 層

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu',
                        input_shape=(150, 150, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(256, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten())
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))

model.summary()

Total params: 6,811,969
Trainable params: 6,811,969
Non-trainable params: 0

Global Average Pooling 層

Flatten層をGlobal Average Pooling層にしたこと以外はすべて同じ

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu',
                        input_shape=(150, 150, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(256, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(GlobalAveragePooling2D())
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))

model.summary()

Total params: 520,513
Trainable params: 520,513
Non-trainable params: 0

結果

・Flatten層 : 6,811,969
・Global Average Pooling層 : 520,513