Tool to reduce the size of Neural Networks models
link to presentation
Clone repository and update python path
repo_name=dietNN # URL of your new repository
username=atinzad # Username for your personal github account
#Clone master
git clone https://github.com/$username/$repo_name
#Or clone a particular branch
my_branch=setup_20180918
git clone -b $my_branch https://github.com/$username/$repo_name
cd $repo_name
echo "export $repo_name=${PWD}" >> ~/.bash_profile
echo "export PYTHONPATH=$repo_name/src:${PYTHONPATH}" >> ~/.bash_profile
source ~/.bash_profile
Create new development branch and switch onto it
branch_name=dev-readme_requisites-20180917 # Name of development branch, of the form 'dev-feature_name-date_of_creation'}}
git checkout -b $branch_name
git push origin $branch_name
- Python 3.6.5
- Tensorflow 1.10.1 (pip install tensorflow #for latest version)
- Numpy 1.14.3 (pip install numpy #for latest version)
- Keras 2.2.2 (pip install keras #for latest version)
- Kerassurgeon 0.1.1 (pip install kerassurgeon #for latest version)
- Optional: GraphViz (sudo apt-get install graphviz)
- Optional: Pydot 1.2.4 (pip install pydot #for latest version)
cd $repo_name
pip install -r requirements.txt
Once done make sure Tensorflow is running as backend (most likely it is) In python, import keras, then go back to shell (this will create keras.json config file)
python
import keras
exit()
Edit $HOME/.keras/keras.json
{
"image_data_format": "channels_last",
"epsilon": 1e-07,
"floatx": "float32",
"backend": "tensorflow"
}
cd ~/dietNN/data/raw
python create_models.py #this will create model.json (in KB range) and model.h5 (in MB range)
example on model.json and model.h5 with reduction request of ~30% in footprint
cd ~/dietNN/src/model
python dietNN.py --m ~/dietNN/data/raw/model.json --w ~/dietNN/data/raw/model.h5 --c 30
Alternativly, using a myconfig.txt file contents of myconfig.txt file
--m=~/dietNN/data/raw/model.json
--w=~/dietNN/data/raw/model.h5
--d=~/dietNN/data/raw/dataset/test
--c=30
Then, in the command prompt
python dietNN.py @myconfig.txt
model_small.json and model_small.h5 will be produced and stored in ~/dietNN/src/model folder
Note that model_small.h5 is ~30% smaller than model.h5
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent