Case studies
Classical Networks
LeNet - 5
input(32,32,1)—-6filters(5x5,s=1)—->conv1(28,28,6)—-avg pool(f=2,s=2)—->pool1(14,14,6)—-16filters(5x5,s=1)—->conv2(10,10,16)—-avg pool(f=2,s=2)—->pool2(5,5,16)—->fully connect(120 neurons)—->fully connect(84 neurons)
AlexNet
VGG - 16
Residual Networks(残差网络)
short cut 穿过Residual Network,有助于解决深度神经网络的梯度消失和梯度爆炸的问题
Network in Network(1x1 convolution)
shrink the channels or increase it or keep it
压缩网络的通道数形成整个网络的瓶颈层,对缩减所需要的计算量有很大帮助
Inception Network
将多种滤波器或者池化(1x1,3x3,5x5,MAX-POOL)的结果堆砌起来
Transfer Learning
下载开源神经网络和预先训练好的权重,只用自己的数据集重新训练最后几层(自己的数据集越多就适合训练越多的神经层)
Data augmentation
- Mirroring
- Random Cropping:随机裁剪
- Rotation
- Shearing
- Local wraping
- Color shifting