深度CNN感受野(Receptive Field)的计算

最新推荐文章于 2024-09-01 08:32:24 发布

XindaBlack

最新推荐文章于 2024-09-01 08:32:24 发布

阅读量619

点赞数

分类专栏：深度学习基础文章标签：深度学习 python

本文链接：https://blog.csdn.net/xindablack/article/details/103913716

版权

深度学习基础专栏收录该内容

2 篇文章 0 订阅

订阅专栏

参考

如何计算感受野(Receptive Field)——原理
 FOMORO AI -> 可视化计算感受野的网站，可以用来验证自己计算的结果

Python代码

这里使用的是从后向前的计算方法，简单且易于理解，也可以直接手算。
$1)*Strade+KernelSize\\ where,KernelSize = KernelSize+(KernelSize-1)(Dilation-1)$

import sys
# VGG16为例
# k_size, stride, padding(1->"same",0->"valid"), dilation
convnet = [[3,1,1,1],[3,1,1,1],[2,2,0,2],[3,1,1,1],[3,1,1,1],[2,2,0,2],[3,1,1,1],[3,1,1,1],[3,1,1,1],[2,2,0,2],[3,1,1,1],[3,1,1,1],[3,1,1,1],[2,2,0,2],
         [3,1,1,1],[3,1,1,1],[3,1,1,1],[2,2,0,2]]
layers_name = ["conv1_1","conv1_2","pool1","conv2_1","conv2_2","pool2","conv3_1","conv3_2","conv3_3","pool3",
             "conv4_1","conv4_2","conv4_3","pool4","conv5_1","conv5_2","conv5_3","pool5"]

vn, vr = [], []
vn.append(224)
vr.append(1)
# l 为当前输出层，从 1 开始
for l in range(1, len(convnet)+1):
    r = 1
    for l_back in range(l, 0, -1):
        k_size = convnet[l_back-1][0]
        s = convnet[l_back-1][1]
        d = convnet[l_back-1][3]
        k_size = k_size + (k_size - 1) * (d - 1)
        if convnet[l_back-1][2] == 1:
            p = (k_size - 1) // 2  # 仅限于 s=1
            # p = (k_size + vn[l_back-1]*(s-1) -s) // 2
        elif convnet[l_back-1][2] == 0:
            p = 0
        else:
            print("Padding is wrong!")
            sys.exit()

        # 计算感受野，从 top 往下层层迭代直到追溯回 input image
        r = (r-1)*s + k_size
        # 计算当前 l 层的输出尺寸
        if l_back == l :
            vn.append((vn[l-1]-k_size+2*p)//s + 1)

    vr.append(r)

print("Input image:")
print("n:", vn[0], " r:", vr[0])
for l in range(1,len(convnet)+1):
    print(layers_name[l-1] + ":")
    print("Output Size:", vn[l], " Kernel Size: ", convnet[l-1][0],
          " Stride: ", convnet[l-1][1], " Receptive Field:", vr[l])

输出结果

Input image:
n: 224  r: 1
conv1_1:
Output Size: 224  Kernel Size:  3  Stride:  1  Receptive Field: 3
conv1_2:
Output Size: 224  Kernel Size:  3  Stride:  1  Receptive Field: 5
pool1:
Output Size: 111  Kernel Size:  2  Stride:  2  Receptive Field: 7
conv2_1:
Output Size: 111  Kernel Size:  3  Stride:  1  Receptive Field: 11
conv2_2:
Output Size: 111  Kernel Size:  3  Stride:  1  Receptive Field: 23
pool2:
Output Size: 55  Kernel Size:  2  Stride:  2  Receptive Field: 27
conv3_1:
Output Size: 55  Kernel Size:  3  Stride:  1  Receptive Field: 35
conv3_2:
Output Size: 55  Kernel Size:  3  Stride:  1  Receptive Field: 43
conv3_3:
Output Size: 55  Kernel Size:  3  Stride:  1  Receptive Field: 51
pool3:
Output Size: 27  Kernel Size:  2  Stride:  2  Receptive Field: 59
conv4_1:
Output Size: 27  Kernel Size:  3  Stride:  1  Receptive Field: 75
conv4_2:
Output Size: 27  Kernel Size:  3  Stride:  1  Receptive Field: 91
conv4_3:
Output Size: 27  Kernel Size:  3  Stride:  1  Receptive Field: 107
pool4:
Output Size: 13  Kernel Size:  2  Stride:  2  Receptive Field: 123
conv5_1:
Output Size: 13  Kernel Size:  3  Stride:  1  Receptive Field: 155
conv5_2:
Output Size: 13  Kernel Size:  3  Stride:  1  Receptive Field: 187
conv5_3:
Output Size: 13  Kernel Size:  3  Stride:  1  Receptive Field: 219
pool5:
Output Size: 6  Kernel Size:  2  Stride:  2  Receptive Field: 251