The project explores the compression techniques for neural nets. Specifically, it takes VGG19 and apply a basic pruning method by trimming weights across the layers in CNN architecture. Additionally, it explores the squeeze net algorithm as a second method by squeezing the layers, improving the activations maps and slimming the frontend filters and input channels so to reduce the size of the network. The inspiration and guidance for squeeze net has been derived from the paper published at this Link
· Python 3
· Pytorch
·Unzip the file CS260finalRB.zip – It unzips into following two folders
Alternatively source Code has been uploaded to github
traincifar10.py # Train the original VGG19 network
vgg.py # module definition
util.py # monitor run time progress
prune.py # Train and test VGG16 with pruning
prune.sh # script to run prune.py with different input thresholds (0.75, 0.5, 0.25, 0.1)
VGGsize.txt - VGG19 original model size
trainlog.txt - log file for original traincifar10.py run
prune_75log.txt, prune_05log.txt, prune_25log.txt # log file with pruning thresholds
train_s.py # Train the original VGG19 network
presqueezewithriginaVGGlog.txt # Log file with original VGG19
vgg_s.py # original module definition
util_s.py # monitor run time progress
vgg_sqz.py # squeeze net with fire layers
train_with_sqz.py # training with squeeze model vgg_sqz
model_after_squeeze_log.txt # Log with squeeze net
Change Directory to VGGPruning
python train_cifar10.py --net vgg # Train the original VGG19 model
Arguments= lr = default 0.1
net = default vgg
r = default resume
python prune.py --net vgg --prune 0.75 # TO Prune the original model with 75% weight reduction
Arguments= prune = default 0.5
net = default vgg
lr = 0.01
./prune.sh -- Shell script to run prune with various thresholds
Change Directory to VGGSqueezenet
python train_s.py # Train the original VGG19 model
python train_with_sqz.py # Apply squeeze algorithm on original model
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent