Keras MaxPooling2D 层不缩小形状

Question

我正在构建一个神经网络，更具体地说是一个 CNN，用于对人体细胞器​​中各种蛋白质的位置进行分类。我在 4 个通道 R、G、B 和 Y 中有 512 x 512 图像：

def ModelMaker(intTuple):

    model = Sequential()

    model.add(Conv2D(8, (7, 7), strides=(2, 2), kernel_initializer='he_normal', bias_initializer='zeros',
                     kernel_regularizer='l2', padding='valid', data_format="channels_last", input_shape=intTuple))

    model.add(Conv2D(16, (3, 3), strides=(1, 1), padding='same', kernel_regularizer='l2'))
    model.add(PReLU(alpha_initializer=VarianceScaling(scale=1.0, mode='fan_in', distribution='normal', seed=None)))
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding='same'))

    model.add(Conv2D(32, (3, 3), strides=(1, 1), padding='same', kernel_regularizer='l2'))
    model.add(PReLU(alpha_initializer=VarianceScaling(scale=1.0, mode='fan_in', distribution='normal', seed=None)))
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding='same'))

    model.add(Conv2D(64, (3, 3), strides=(1, 1), padding='same', kernel_regularizer='l2'))
    model.add(PReLU(alpha_initializer=VarianceScaling(scale=1.0, mode='fan_in', distribution='normal', seed=None)))
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding='same'))

    model.add(Conv2D(128, (3, 3), strides=(1, 1), padding='same', kernel_regularizer='l2'))
    model.add(PReLU(alpha_initializer=VarianceScaling(scale=1.0, mode='fan_in', distribution='normal', seed=None)))
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding='same'))

    model.add(Conv2D(256, (3, 3), strides=(1, 1), padding='same', kernel_regularizer='l2'))
    model.add(PReLU(alpha_initializer=VarianceScaling(scale=1.0, mode='fan_in', distribution='normal', seed=None)))
    model.add(AveragePooling2D(pool_size=(2, 2), strides=(1, 1), padding='same'))

    model.add(Flatten())

    model.add(Dense(28))
    model.add(Activation('softmax'))

    return model

model = ModelMaker((512, 512, 4))
model.summary()

一切都很好，但是当我创建模型并运行model.summary() 并遍历各个层时，有些奇怪。对于每个卷积，我有以下层序列：Conv2D ---> PReLU ---> MaxPooling2D/AveragePooling2D。但是，模型摘要如下所示：

Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_6 (Conv2D)            (None, 253, 253, 8)       1576      
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 253, 253, 16)      1168      
_________________________________________________________________
p_re_lu_5 (PReLU)            (None, 253, 253, 16)      1024144   
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 253, 253, 16)      0         
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 253, 253, 32)      4640      
_________________________________________________________________
p_re_lu_6 (PReLU)            (None, 253, 253, 32)      2048288   
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 (None, 253, 253, 32)      0         
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 253, 253, 64)      18496     
_________________________________________________________________
p_re_lu_7 (PReLU)            (None, 253, 253, 64)      4096576   
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 253, 253, 64)      0         
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 253, 253, 128)     73856     
_________________________________________________________________
p_re_lu_8 (PReLU)            (None, 253, 253, 128)     8193152   
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 253, 253, 128)     0         
_________________________________________________________________
conv2d_11 (Conv2D)           (None, 253, 253, 256)     295168    
_________________________________________________________________
p_re_lu_9 (PReLU)            (None, 253, 253, 256)     16386304  
_________________________________________________________________
average_pooling2d_1 (Average (None, 253, 253, 256)     0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 16386304)          0         
_________________________________________________________________
dense_1 (Dense)              (None, 28)                458816540 
_________________________________________________________________
activation_1 (Activation)    (None, 28)                0         
=================================================================
Total params: 490,959,908
Trainable params: 490,959,908
Non-trainable params: 0
_________________________________________________________________

降维似乎有问题。对于前两个维度（253、253、XX），张量的大小都是相同的。它保持在 253 不变......只有通道大小增加，这是正常的。我删除了最后两个密集层堆栈，因为模型甚至不会将自己定义为运行时崩溃，因为内存已用完。我猜这是由于当形状（253、253、256）变平并通过时，致密层的尺寸非常大。整个模型不适合 RAM。救命！

Answer 1

使用strides=(2, 2) 或将其留空。然后，max_pooling2d 将起作用。可训练参数的数量将显着减少。如果未指定，max_pooling2d 将默认为 pool_size。请参阅docs。