SPP模块主要是为了应对当网络中存在全连接层时,对中间层级的输出特征的形状具有不变的要求,其具体思想可以理解为将具有一定形状的输出特征分成指定数量的子特征,然后使用池化的手段将其转化成具有一定维度的特征矩阵,最后进行拼接,从而使网络具有输入图片任意尺寸的“适应能力”。
具体实现代码如下(基于pytorch):
import math
import torch
import torch.nn.functional as F
class SPPnet(torch.nn.Module):
def __init__(self, num_layers, pool_type='max'):
super(SPPnet, self).__init__()
self.num_level = num_layers
self.pool_type = pool_type
def forward(self,input_feature):
batch, c, h, w = input_feature.size()
for i in range(self.num_level):
level = 1+1
kernel_size = (math.ceil(h/level), math.ceil(w/level))
stride = (math.ceil(h/level), math.ceil(w/level))
padding = (math.floor((kernel_size[0]* level-h+1)/2), math.floor((kernel_size[1]* level-w+1)/2))
if self.pool_type =='max':
tensor = F.max_pool2d(input_feature, kernel_size = kernel_size, padding = padding).view(batch,-1)
else:
tensor = F.avg_pool2d(input_feature, kernel_size = kernel_size, padding = padding).view(batch,-1)
#然后就可以按照SPPnet后面所接的全连接层Fc对tensor进行相应的操作了。