1. 阅读论文:Understanding the Effective Receptive Field in Deep Convolutional Neural Networks 理解感受野 定义:receptive field, or field of view (感受野) A unit in convolutional networks only depends on a region of the input. This region in the input is the recepti
CNN的权值正交性和特征正交性,在一定程度上是和特征表达的差异性存在一定联系的. 下面两篇论文,一篇是在训练中对权值添加正交正则提高训练稳定性,一篇是对特征添加正交性的损失抑制过拟合. 第一篇:Orthonormality Regularization Xie D, Xiong J, Pu S. All You Need is Beyond a Good Init: Exploring Better Solution for Training Extremely Deep Convolution
0 - inplace 在pytorch中,nn.ReLU(inplace=True)和nn.LeakyReLU(inplace=True)中存在inplace字段.该参数的inplace=True的意思是进行原地操作,例如: x=x+5是对x的原地操作 y=x+5,x=y不是对x的原地操作 所以,如果指定inplace=True,则对于上层网络传递下来的tensor直接进行修改,可以少存储变量y,节省运算内存. inplace=True means that it will modify th