Question

我正在尝试了解GoogleNet的接受领域。

我使用了如下所示的程序。

import math
convnet =   [[7,2,3],[1,1,0],[3,2,0],[1,1,0],[1,1,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[3,2,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[3,2,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[5,3,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[5,3,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[3,2,1],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[1,1,0],[1,1,0],[3,1,1],[1,1,0],[7,1,1]]
layer_names = ["conv1/7x7_s2","conv1/relu_7x7","pool1/3x3_s2","pool1/norm1","conv2/3x3_reduce","conv2/relu_3x3_reduce","conv2/3x3","conv2/relu_3x3","pool2/3x3_s2","inception_3a/3x3_reduce","inception_3a/relu_3x3_reduce","inception_3a/3x3","inception_3a/relu_3x3","inception_3b/3x3_reduce","inception_3b/relu_3x3_reduce","inception_3b/3x3","inception_3b/relu_3x3","pool3/3x3_s2","inception_4a/3x3_reduce","inception_4a/relu_3x3_reduce","inception_4a/3x3","inception_4a/relu_3x3","loss1/ave_pool","inception_4b/3x3_reduce","inception_4b/relu_3x3_reduce","inception_4b/3x3","inception_4b/relu_3x3","inception_4c/3x3_reduce","inception_4c/relu_3x3_reduce","inception_4c/3x3","inception_4c/relu_3x3","inception_4d/3x3_reduce","inception_4d/relu_3x3_reduce","inception_4d/3x3","inception_4d/relu_3x3","loss2/ave_pool","inception_4e/3x3_reduce","inception_4e/relu_3x3_reduce","inception_4e/3x3","inception_4e/relu_3x3","pool4/3x3_s2","inception_5a/3x3_reduce","inception_5a/relu_3x3_reduce","inception_5a/3x3","inception_5a/relu_3x3","inception_5b/3x3_reduce","inception_5b/relu_3x3_reduce","inception_5b/3x3","inception_5b/relu_3x3","pool5/7x7_s1"]
imsize = 720
def outFromIn(isz, layernum, net = convnet):
    if layernum>len(net): layernum=len(net)

    totstride = 1
    insize = isz
    #for layerparams in net:
    for layer in range(layernum):
        fsize, stride, pad = net[layer]
        outsize = (insize - fsize + 2*pad) / stride + 1
        insize = outsize
        totstride = totstride * stride
    return outsize, totstride

def inFromOut( layernum, net = convnet):
    if layernum>len(net): layernum=len(net)
    outsize = 1
    #for layerparams in net:
    for layer in reversed(range(layernum)):
        fsize, stride, pad = net[layer]
        outsize = ((outsize -1)* stride) + fsize
    RFsize = outsize
    return RFsize

if __name__ == '__main__':

    print "layer output sizes given image = %dx%d" % (imsize, imsize)
    for i in range(len(convnet)):
        p = outFromIn(imsize,i+1)
        rf = inFromOut(i+1)
        print "Layer Name = %s, Output size = %3d, Stride = % 3d, RF size = %3d" % (layer_names[i], p[0], p[1], rf)

每层的感受野计算为

layer output sizes given image = 224x224
Layer Name = conv1/7x7_s2, Output size = 112, Stride =   2, RF size =   7
Layer Name = conv1/relu_7x7, Output size = 112, Stride =   2, RF size =   7
Layer Name = pool1/3x3_s2, Output size =  55, Stride =   4, RF size =  11
Layer Name = pool1/norm1, Output size =  55, Stride =   4, RF size =  11
Layer Name = conv2/3x3_reduce, Output size =  55, Stride =   4, RF size =  11
Layer Name = conv2/relu_3x3_reduce, Output size =  55, Stride =   4, RF size =  11
Layer Name = conv2/3x3, Output size =  55, Stride =   4, RF size =  19
Layer Name = conv2/relu_3x3, Output size =  55, Stride =   4, RF size =  19
Layer Name = pool2/3x3_s2, Output size =  27, Stride =   8, RF size =  27
Layer Name = inception_3a/3x3_reduce, Output size =  27, Stride =   8, RF size =  27
Layer Name = inception_3a/relu_3x3_reduce, Output size =  27, Stride =   8, RF size =  27
Layer Name = inception_3a/3x3, Output size =  27, Stride =   8, RF size =  43
Layer Name = inception_3a/relu_3x3, Output size =  27, Stride =   8, RF size =  43
Layer Name = inception_3b/3x3_reduce, Output size =  27, Stride =   8, RF size =  43
Layer Name = inception_3b/relu_3x3_reduce, Output size =  27, Stride =   8, RF size =  43
Layer Name = inception_3b/3x3, Output size =  27, Stride =   8, RF size =  59
Layer Name = inception_3b/relu_3x3, Output size =  27, Stride =   8, RF size =  59
Layer Name = pool3/3x3_s2, Output size =  13, Stride =  16, RF size =  75
Layer Name = inception_4a/3x3_reduce, Output size =  13, Stride =  16, RF size =  75
Layer Name = inception_4a/relu_3x3_reduce, Output size =  13, Stride =  16, RF size =  75
Layer Name = inception_4a/3x3, Output size =  13, Stride =  16, RF size = 107
Layer Name = inception_4a/relu_3x3, Output size =  13, Stride =  16, RF size = 107
Layer Name = inception_4b/3x3_reduce, Output size =  13, Stride =  16, RF size = 107
Layer Name = inception_4b/relu_3x3_reduce, Output size =  13, Stride =  16, RF size = 107
Layer Name = inception_4b/3x3, Output size =  13, Stride =  16, RF size = 139
Layer Name = inception_4b/relu_3x3, Output size =  13, Stride =  16, RF size = 139
Layer Name = inception_4c/3x3_reduce, Output size =  13, Stride =  16, RF size = 139
Layer Name = inception_4c/relu_3x3_reduce, Output size =  13, Stride =  16, RF size = 139
Layer Name = inception_4c/3x3, Output size =  13, Stride =  16, RF size = 171
Layer Name = inception_4c/relu_3x3, Output size =  13, Stride =  16, RF size = 171
Layer Name = inception_4d/3x3_reduce, Output size =  13, Stride =  16, RF size = 171
Layer Name = inception_4d/relu_3x3_reduce, Output size =  13, Stride =  16, RF size = 171
Layer Name = inception_4d/3x3, Output size =  13, Stride =  16, RF size = 203
Layer Name = inception_4d/relu_3x3, Output size =  13, Stride =  16, RF size = 203
Layer Name = inception_4e/3x3_reduce, Output size =  13, Stride =  16, RF size = 203
Layer Name = inception_4e/relu_3x3_reduce, Output size =  13, Stride =  16, RF size = 203
Layer Name = inception_4e/3x3, Output size =  13, Stride =  16, RF size = 235
Layer Name = inception_4e/relu_3x3, Output size =  13, Stride =  16, RF size = 235
Layer Name = pool4/3x3_s2, Output size =   7, Stride =  32, RF size = 267
Layer Name = inception_5a/3x3_reduce, Output size =   7, Stride =  32, RF size = 267
Layer Name = inception_5a/relu_3x3_reduce, Output size =   7, Stride =  32, RF size = 267
Layer Name = inception_5a/3x3, Output size =   7, Stride =  32, RF size = 331
Layer Name = inception_5a/relu_3x3, Output size =   7, Stride =  32, RF size = 331
Layer Name = inception_5b/3x3_reduce, Output size =   7, Stride =  32, RF size = 331
Layer Name = inception_5b/relu_3x3_reduce, Output size =   7, Stride =  32, RF size = 331
Layer Name = inception_5b/3x3, Output size =   7, Stride =  32, RF size = 395
Layer Name = inception_5b/relu_3x3, Output size =   7, Stride =  32, RF size = 395
Layer Name = pool5/7x7_s1, Output size =   3, Stride =  32, RF size = 587

我不明白的是，对于输入图像大小为224，最后一层的感知字段大小应为224左右。

对于pool5 / 7x7_s1图层，现在是587。

初始模块inception_4e的感知字段大小与输入图像大小224的大小相同。

我的计算有什么问题？在论文中，它清楚地表明，RGB颜色空间中的感受野是224 x 224，平均值为零。

Answer 1

接收字段与输入大小无关：
感受野的大小告诉您“像素”在计算网络的单个输出时会影响（或参与）请注意，在计算感知字段时，您完全忽略填充（！）。

另一方面，当您想要计算网络的输出大小时，您会考虑内核大小和步幅（影响感知字段），但也输入大小和填充。

因此，感受野大小和输入/输出大小是相关的，但决不相同。

receptive field arithmetics的精彩指南。

Google Net

1 个答案: