Classic-CNN-

This is a description of classic convolutional neural network and its implementation in Pytorch
We briefly introduce the structure and #params and FLOPs of these network

1. LeNet

LeNet, proposed in 1998 by LeCun, contains 7 layers including 2 conv, 2 pooling and 3 fc layers

2. AlexNet

AlexNet, proposed in 2012, contains 8 layers including 5 conv and 3 fc layers.

tricks

ReLU
Dropout
Local Response Normalization
Overlapping Pooling

3. VGG16

VGG, proposed in 2014, is a series of neural network and contains 16-19 layers.
We only implement VGG16 which consists of 13 conv, 5 pooling and 3 fc layers

tricks

Deeper
smaller kernel_size

4. SqueezeNet

SqueezeNet, proposed in 2016, is a lightweight neural network.
(reimplementation)
Total param:1,244,448
MAdd: 1.67G
Flops: 838.94M

tricks

1. Replace 3x3 filters with 1x1 filters
   
2. Decrease the number of input channels to 3x3 filters
   
3. Downsample late in the network so that convolution layers have large activation maps