卷积网络感受野计算

import copy
####top_to_bottom calculate
def cal_receptive_filed(kernel_stride):
    fix_kernel_stride = copy.deepcopy(kernel_stride)
    layer_receptive_field = []
    for j in range(0,-len(fix_kernel_stride),-1):
        if j == 0:
            kernel_stride= fix_kernel_stride
        else:
            kernel_stride = fix_kernel_stride[:j]
        for i in range(len(kernel_stride)-1,-1,-1):
            if i == len(kernel_stride)-1:
                RF = kernel_stride[-1][0]
            else:
                RF = (RF-1)*kernel_stride[i][1] +kernel_stride[i][0]
        layer_receptive_field.append(RF)
    return layer_receptive_field
if __name__ == "__main__":
    kernel_stride = [[3,1],
                     [3, 1],
                     [2,2],  ###//2
                     [3,1],
                     [3, 1],
                     [2,2], ###//4
                     [3,1],
                     [3, 1],
                     [2,2], ###//8
                     [3, 1],
                     [3, 1],
                     [2,2], ##//16
                     [3, 1],
                     [3, 1],
                     [2,2],###//32
                     [3, 1],
                     [3, 1],
                     [2,2],###//64
                     [3, 1],
                     [3, 1],
                     [3, 1],
                     [3, 1],
                     ]
    RF = cal_receptive_filed(kernel_stride)
    print zip(kernel_stride,RF[::-1])

对于检测任务,通常需要计算每一层的感受野去安排每一级的预测目标.具体原理请参考他人文章https://blog.csdn.net/program_developer/article/details/80958716

有一个网站Fomoro AI可以计算每一层的感受野,和上面自己写的代码是一样的结果,仅供参考.