Awesome CNNs

This repo contains a collection of CNN network structures.

For each kind of CNN network, we provide a pytorch implementation, a caffe implementation for NetScope visualization, and additional pictures(mostly from original paper) to help you understand how it works. Links to original paper and my notes(in Chinese) are also provided.

Furthermore, in experiments folder, there is a training script for you to train each kind of network with Pytorch on dataset CIFAR-10.

Usage

1.Clone this repo.

git clone https://github.com/ddlee96/awesome_cnn.git

2.Install PyTorch and TorchVision following instructions on PyTorch website.

(Optional)Install dependencies using pipenv:

pip install pipenv
cd awesome_cnn
pipenv install

pipenv will install dependencies based on Pipfile. You can also modify this Pipfile to match your environment.

3.Prepare dataset

cd awesome_cnn
mkdir data

Then just run python experiments/experiment.py --data cifar, it will create data/cifar/ and download cifar-10 dataset for you.

(Optional)You can also use dataset MNIST and Fashion-MNIST, it requires Pillow package. Install it using pip install Pillow.

4.Start Training

python experiments/experiment.py --model resnet50 --data cifar --batch_size 182 --gpus 0,1,2,3

It will train model ResNet50 on cifar10 with batch size 128. More options can be find in experiment.py.

CNN Network List

Network Name	PaperNotes	Paper	Netscope	Proposed Date
MobileNets	MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications	arxiv	-	17.04
ResNeXt	Aggregated Residual Transformations for Deep Neural Networks	arxiv	Visualization	16.11
Xception	Xception: Deep Learning with Depthwise Seperable Convolutions	arxiv	Visualization	16.10
DenseNet	Densely Connected Convolutional Networks	arxiv	Visualization	16.08
ResNet-v2	Identity Mappings in Deep Residual Networks	arxiv	-	16.03
Inception-ResNet	Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning	arxiv	Visualization	16.02
SqueezeNet	SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size	arxiv	-	16.02
Inception-v3	Rethinking the Inception Architecture for Computer Vision	arxiv	Visualization	15.12
ResNet	Deep Residual Learning for Image Recognition	arxiv	Visualization	15.12
GoogLeNet	Going deeper with convolutions	arxiv	Visualization	14.09
VGG	Very Deep Convolutional Networks for Large-Scale Image Recognition	arxiv	Visualization	14.09
AlexNet	ImageNet Classification with Deep ConvolutionalNeural Networks	NIPS	Visualization	-
LeNet	Gradient-based learning applied to document recognition	Lecun's Homepage	-	1998

To be added

• SqueezeNet
• ResNet-v2
• SENet
• LeNet
• DPN
• NASNet
• MobileNet V2