inception_v4.py 文件源码

def reduction_A(input):
    if K.image_dim_ordering() == "th":
        channel_axis = 1
    else:
        channel_axis = -1

    r1 = conv_block(input, 384, 3, 3, subsample=(2, 2), border_mode='valid')

    r2 = conv_block(input, 192, 1, 1)
    r2 = conv_block(r2, 224, 3, 3)
    r2 = conv_block(r2, 256, 3, 3, subsample=(2, 2), border_mode='valid')

    r3 = MaxPooling2D((3, 3), strides=(2, 2), border_mode='valid')(input)

    m = merge([r1, r2, r3], mode='concat', concat_axis=channel_axis)
    return m