res_net.py

def res_net(shape):
    from keras.models import Model
    from keras.layers import Dense, Conv2D, MaxPooling2D, Flatten, Input, BatchNormalization, Add, Activation, GlobalAveragePooling2D
    from keras.regularizers import l2

    def resblock(filters, kernel_size=(3, 3), increase=False):
        strides = (2, 2) if increase else (1, 1)
        def _res_block(x):
            x_ = Conv2D(filters, kernel_size,
                strides=strides,
                padding='same',
                kernel_regularizer=l2(weight_decay),
                activation='relu')(x)
            x_ = BatchNormalization()(x_)
            x_ = Conv2D(filters, kernel_size,
                strides=(1, 1),
                padding='same',
                kernel_regularizer=l2(weight_decay),
                activation='relu')(x_)
            if increase:
              x = Conv2D(filters, (1, 1),
                  strides=(2, 2),
                  padding='same',
                  kernel_regularizer=l2(weight_decay),
                  activation='relu')(x)
            x = Add()([x_, x])
            x = BatchNormalization()(x)
            x = Activation('relu')(x)
            return x
        return _res_block

    weight_decay = 1e-4
    inputs = Input(shape=shape)

    # 32 * 32 * 3
    x = Conv2D(64, (7, 7), padding='same', kernel_regularizer=l2(weight_decay), activation='relu')(inputs)
    x = BatchNormalization()(x)
    x = MaxPooling2D(pool_size=(2, 2))(x)

    # 16 * 16 * 64 → 16 * 16 * 64
    x = resblock(64, increase=True)(x)
    x = resblock(64)(x)
    x = resblock(64)(x)

    x = MaxPooling2D(pool_size=(2, 2))(x)

    # 8 * 8 * 64 → 8 * 8 * 128
    x = resblock(128, increase=True)(x)
    x = resblock(128)(x)
    x = resblock(128)(x)
    x = resblock(128)(x)

    x = MaxPooling2D(pool_size=(2, 2))(x)

    # 4 * 4 * 128 → 4 * 4 * 256
    x = resblock(256, increase=True)(x)
    x = resblock(256)(x)
    x = resblock(256)(x)
    x = resblock(256)(x)
    x = resblock(256)(x)
    x = resblock(256)(x)

    x = MaxPooling2D(pool_size=(2, 2))(x)

    # 2 * 2 * 256 → 2 * 2 * 512
    x = resblock(512, increase=True)(x)
    x = resblock(512)(x)
    x = resblock(512)(x)

    x = GlobalAveragePooling2D()(x)
    #  512 → 10
    y = Dense(10, activation='softmax')(x)

    model = Model(inputs=inputs, outputs=y)
    model.compile(
        loss='categorical_crossentropy',
        optimizer='adam',
        metrics=['accuracy']
    )

    return model