为什么使用3x3的小的卷积核,是否能够达到5x5, 7x7,11x11卷积核的效果
- 之所以使用3x3,而非2x2,1x1的核,是因为(3x3)是捕获图像中左/右,上/下,中心概念的最小尺寸
- 两层3x3的卷积层的叠加的感受野,相当于一个5x5的卷积的感受野;三层3x3的卷积的感受野,相当于一个7x7的卷积的感受野(限定条件,每个卷积具有相同的padding = [1,1], stride = [1,1]),感受野的计算公式,见下节
- 有参数量的计算公式可知,两个3x3的卷积本身的参数量为3x3x2 = 18,一个5x5卷积的参数量为25;三个卷积3x3卷积的参数量为27,一个7x7卷积的参数量为49
感受野
感受野的计算公式
该感受野计算网站实现了对每层卷积层计算感受野大小。
计算公式定义如下:
参数 | 意义 | 英文 |
---|---|---|
n | feature map的数量 | num of features |
p | 卷积核尺寸 | convolution kernel size |
k | 卷积核padding尺寸 | convolution padding size |
s | 卷积核stride尺寸 | convolution stride size |
r | 感受野的大小 | receptive field size |
j | 两个相邻特征值之间的跨度 | jump(distance between two consecutive features) |
start | 特征图中第一个特征的在原图上中心坐标 | center coordinate of the first feature |
feature map输出尺寸公式:
n
o
=
n
i
+
2
p
−
k
s
n_o=\frac{n_i+2p-k}{s}
no=sni+2p−k
特征图跨度公式:
j
o
=
j
i
×
s
j_o = j_i\times s
jo=ji×s
感受野公式:
r
o
=
r
i
+
(
k
−
1
)
×
j
i
r_o=r_i+(k-1)\times j_i
ro=ri+(k−1)×ji
我们来计算下两个3x3的卷积,和5x5的卷积,在输入图片input_size = 224, p = 1, s = 1的情况下的感受野输出
初始值:
#对输入图片,由以下值,即初始的感受野大小和特征之间的跨度都为1
imsize = 224
j_i = 1
r_i = 1
两个3x3卷积的感受野大小:
- 第一个3x3卷积的感受野大小
r o = r i + ( k − 1 ) × j i = 1 + ( 3 − 1 ) × 1 = 3 r_o =r_i+(k-1)\times j_i =1 + (3-1)\times 1 =3 ro=ri+(k−1)×ji=1+(3−1)×1=3
- 第二个3x3卷积的感受野输出
r o = r i + ( k − 1 ) × j i = 3 + ( 3 − 1 ) × 1 = 5 r_o =r_i+(k-1)\times j_i =3 + (3-1)\times 1 =5 ro=ri+(k−1)×ji=3+(3−1)×1=5
单个5x5卷积的感受野大小:
r o = r i + ( k − 1 ) × j i = 1 + ( 5 − 1 ) × 1 = 5 r_o =r_i+(k-1)\times j_i =1 + (5-1)\times 1 =5 ro=ri+(k−1)×ji=1+(5−1)×1=5
由此可见,在步长stride和padding一致的条件下,两个3x3的卷积和一个5x5的卷积,感受野是一样的
同样可以计算3个3x3的卷积的感受野大小为7,同一个7x7的卷积的感受野相同
感兴趣的区域计算方式
每层卷积处理后,会输出对应的feature_map,每个fature_map上每一个特征点会对应一个感受野大小,那这个感受野区域在原图的位置,又是怎么计算的?
感受野的大小我们已经知道,如果再求出该感受野区域在原图中映射的中心点,就可以以求出感受野区域了
计算公式定义如下:
参数 | 意义 | 英文 |
---|---|---|
start | 特征图中第一个特征的在原图上中心坐标 | center coordinate of the first feature |
注意,以上start是指每个特征图中,第一个特征点在原图中的感受野区域的中心点,计算公式为:
s t a r t o = s t a r t i + ( k − 1 2 − p ) × j i start_o = start_i + (\frac{k-1}{2}-p)\times j_i starto=starti+(2k−1−p)×ji
还是以两个3x3的卷积计算为例,计算下每层处理后,每个特征图的第一个特征点在原图上的对应位置:
初始值:
#对输入图片,由以下值,即初始的感受野大小和特征之间的跨度都为1
imsize = 224
j_i = 1
r_i = 1
start = 0.5
第一层卷积处理后的,feature_map大小为3x3,第一个feature点的start为:
s
t
a
r
t
o
=
s
t
a
r
t
i
+
(
k
−
1
2
−
p
)
×
j
i
=
0.5
+
+
(
3
−
1
2
−
1
)
×
1
=
0.5
start_o = start_i + (\frac{k-1}{2}-p)\times j_i = 0.5 + + (\frac{3-1}{2}-1)\times 1 =0.5
starto=starti+(2k−1−p)×ji=0.5++(23−1−1)×1=0.5
j o = j i × s = 1 j_o = j_i\times s = 1 jo=ji×s=1
同理,第二层处理后,feature_map大小为5x5,第一个feature点的start为:
s t a r t o = s t a r t i + ( k − 1 2 − p ) × j i = 0.5 + + ( 3 − 1 2 − 1 ) × 1 = 0.5 start_o = start_i + (\frac{k-1}{2}-p)\times j_i = 0.5 + + (\frac{3-1}{2}-1)\times 1 =0.5 starto=starti+(2k−1−p)×ji=0.5++(23−1−1)×1=0.5
j o = j i × s = 1 j_o = j_i\times s = 1 jo=ji×s=1
如果,现在要计算,第二层输出的feature_map中,第[2,2]个特征点对应的中心点坐标,使用以下公示计算
其他特征点对应的区域,按照以下公示进行转换
start_i_x
= start_0-1_x + feature_i_x * j
= 0.5 + 2*1 = 2.5
start_i_x
= start_0_y + feature_i_y * j
= 0.5 + 2*1 = 2.5
其中,start_0表示第当前特征图一个特征点在原图中的感受野的中心位置,feature_i表示当前特征图中,第i个特征点的点位,j表示当前特征图相邻特征点之间的跨度
感受野计算实现代码
# [filter size, stride, padding]
#Assume the two dimensions are the same
#Each kernel requires the following parameters:
# - k_i: kernel size
# - s_i: stride
# - p_i: padding (if padding is uneven, right padding will higher than left padding; "SAME" option in tensorflow)
#
#Each layer i requires the following parameters to be fully represented:
# - n_i: number of feature (data layer has n_1 = imagesize )
# - j_i: distance (projected to image pixel distance) between center of two adjacent features
# - r_i: receptive field of a feature in layer i
# - start_i: position of the first feature's receptive field in layer i (idx start from 0, negative means the center fall into padding)
import math
def outFromIn(conv, layerIn):
n_in = layerIn[0]
j_in = layerIn[1]
r_in = layerIn[2]
start_in = layerIn[3]
k = conv[0]
s = conv[1]
p = conv[2]
n_out = math.floor((n_in - k + 2*p)/s) + 1
actualP = (n_out-1)*s - n_in + k
pR = math.ceil(actualP/2)
pL = math.floor(actualP/2)
j_out = j_in * s
r_out = r_in + (k - 1)*j_in
start_out = start_in + ((k-1)/2 - pL)*j_in
return n_out, j_out, r_out, start_out
def printLayer(layer, layer_name):
print(layer_name + ":")
print("\t n features: %s \n \t jump: %s \n \t receptive size: %s \t start: %s " % (layer[0], layer[1], layer[2], layer[3]))
def showReceptiveField(currentLayer, convnet, layer_names):
layerInfos = []
printLayer(currentLayer, "input image")
for i in range(len(convnet)):
currentLayer = outFromIn(convnet[i], currentLayer)
layerInfos.append(currentLayer)
printLayer(currentLayer, layer_names[i])
print ("------------------------")
layer_name = input ("Layer name where the feature in: ") # 这里要修改
layer_idx = layer_names.index(layer_name)
idx_x = int(input ("index of the feature in x dimension (from 0)")) # 这里要修改
idx_y = int(input ("index of the feature in y dimension (from 0)")) # 这里要修改
n = layerInfos[layer_idx][0]
j = layerInfos[layer_idx][1]
r = layerInfos[layer_idx][2]
start = layerInfos[layer_idx][3]
assert(idx_x < n)
assert(idx_y < n)
print ("receptive field: (%s, %s)" % (r, r))
print ("center: (%s, %s)" % (start+idx_x*j, start+idx_y*j))
if __name__ == '__main__':
#first layer is the data layer (image) with n_0 = image size; j_0 = 1; r_0 = 1; and start_0 = 0.5
imsize = 224
currentLayer = [imsize, 1, 1, 0.5]
print ("-------Net summary 3x3 ------")
convnet_3x3 = [[3,1,1],[3,1,1]]
layer_names_3x3 = ['conv1','conv2']
showReceptiveField(currentLayer, convnet_3x3,layer_names_3x3)
print ("-------Net summary 5x5 ------")
convnet_5x5 = [[5,1,1]]
layer_names_5x5 = ['conv1']
showReceptiveField(currentLayer, convnet_5x5,layer_names_5x5)
convnet = [[11,4,0],[3,2,0],[5,1,2],[3,2,0],[3,1,1],[3,1,1],[3,1,1],[3,2,0],[6,1,0], [1, 1, 0]]
layer_names = ['conv1','pool1','conv2','pool2','conv3','conv4','conv5','pool5','fc6-conv', 'fc7-conv']
showReceptiveField(currentLayer, convnet,layer_names)
两层3x3卷积,在224的输入上的计算结果:
-------Net summary 3x3 ------
input image:
n features: 224
jump: 1
receptive size: 1 start: 0.5
conv1:
n features: 224
jump: 1
receptive size: 3 start: 0.5
conv2:
n features: 224
jump: 1
receptive size: 5 start: 0.5
------------------------
Layer name where the feature in: conv2
index of the feature in x dimension (from 0)2
index of the feature in y dimension (from 0)2
receptive field: (5, 5)
center: (2.5, 2.5)
相关的理论技术请参考网站A guide to receptive field arithmetic for Convolutional Neural Networks