EdgeLearning

EdgeLearning is a Deep Learning library for Edge Computing developed for high performance, low footprint on embedded devices and distributed systems.

The project library is supported on MacOS, Windows and Linux systems, with Clang and GNU GCC compiler.

Fast setup

Commands to build the library with default configuration:

git submodule update --init --recursive
mkdir build && cd build
cmake ..
cmake --build .

Commands to build the library with default configuration and Python interface:

git submodule update --init --recursive
mkdir build && cd build
cmake .. -DBUILD_PYTHON=1
cmake --build .

Open the example folder to get started and learn how to use the EdgeLearning library on known datasets.

Dependencies

Clone the Git repository:

git clone https://github.com/mircodemarchi/EdgeLearning.git EdgeLearning
cd EdgeLearning

The Git repository has submodules, in order to download the submodules, it is necessary to issue:

git submodule update --init --recursive

Building

Setup the build subdirectory:

mkdir build && cd build

Then you need to prepare the build environment. Choose your desired configuration with the following parameters:

With GNU GCC compiler:

cmake -DCMAKE_C_COMPILER=gcc-11 -DCMAKE_CXX_COMPILER=g++-11 ..

With Clang compiler:

cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ ..

Enable MLPACK: the high level api of EdgeLearning allows to select Mlpack functionalities statically or dynamically.
```
cmake -DENABLE_MLPACK=1 ..
```
N.B. The option ENABLE_MLPACK is not supported in MacOS with GNU GCC compiler.
Build with Python interface:
```
cmake -DBUILD_PYTHON=1 ..
```

Build and install the library:

make -j 
sudo make install

Profiling

The profiling routines included in EdgeLearning explore different types of datasets by varying architecture and model parameters.

There are 4 profiling routines:

Classification problem solved with Feedforward Neural Network: profile_fnn_classification;
Regression problem solved with Feedforward Neural Network: profile_fnn_regression;
Classification problem solved with Recurrent Neural Network: Work in Progress;
Regression problem solved with Recurrent Neural Network: Work in Progress;

In order to execute the routines of profiling you need to compile the library as explained in Building section. Then you can execute the profiling routines as following:

./profile/profile_fnn_classification
./profile/profile_fnn_regression

The profiling routines will create a folder of the following pattern: profile__, where framework could be "edgelearning" or "mlpack" based on the library configuration choice of the ENABLE_MLPACK field, and could be "fnn_classification", "fnn_regression", "rnn_classification" or "rnn_regression", based on the profiling routine executed. The folder will contain all the performance measure produced by the execution of the profiling routine divided in files that represent the model architecture and parameter used in the profiling steps.

Contribution guidelines

If you would like to contribute to Edge Learning, please contact the developer:

Mirco De Marchi mirco.demarchi@univr.it
Luca Geretti luca.geretti@univr.it