CNN inference on CPU using frugally-deep

The aim of this project is to perform inference of a CNN on CPU using the frugally-deep library. Thus, it is possible to train a model using Keras or Tensorflow in python, save it in h5 format and then convert it into a format usable by frugally-deep in order to perform the prediction directly from a c++ program.

Setting up the environment

Verify your cmake version

To install frugally-deep, you need cmake 3.14 or higher. You can check your cmake version by executing:

cmake --version

If your version of cmake is too old, you should upgrade it (see https://cmake.org/download/).

Install cnpy

The library cnpy allows you to open and load .npy files (numpy) in a c++ program. To install it properly, you can run the following commands in a shell:

sudo apt-get install zlib1g zlib1g-dev
git clone https://github.com/rogersce/cnpy
cd cnpy/
mkdir build
cd build/
cmake ..
make
sudo make install

Install tqdm

In order to facilitate the visualisation of the prediction progress, we use a c++ version of tqdm. This is a header-only. You just have to copy the header to the right place by running the following command:

cd lib
sudo mkdir /usr/local/include/tqdm
sudo cp tqdm.h /usr/local/include/tqdm

Install frugally-deep

Now you just have to run the installation script. It will take care of cloning the right directories and copying the files to be included in the right places.

cd tools
chmod +x setup.sh
sudo ./setup.sh

Check your LD_LIBRARY_PATH

Be sure that /usr/local/lib is in the LD_LIBRARY_PATH environment variable. You can achieve this automatically by adding this line at the end of your ~.profile file:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib

Testing the environment

Go to the root of the directory, create a build folder and simply do:

make