YANN is the Neural Network library implemented from ground up using the Eigen library for matrix operations. The code uses many C++17 features (it was compiled and tested using g++ 8.3 and should compile on most modern compilers).
- Fully connected networks, CNN, RNN, LSTM, ...
- Gradient descent, AdaGrad, AdaDelta, ...
- Sigmoid, Tanh, Quadratic Cost, Entropy Cost, ...
Install Boost and Eigen, for example on Ubuntu:
sudo apt install libboost-all-dev libeigen3-dev zlib1g-dev
Optionally, install Intel MKL (see Eigen docs for more details):
cd /tmp
wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB
sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB
sudo sh -c 'echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list'
sudo apt-get update && sudo apt-get install intel-mkl
Followed by the usual
./configure && make && make check
The two helper scripts configure-debug.sh and configure-release.sh
could be used instead to pass reasonable debug and optimization flags
to g++.
The YANN library supports sequential, parallel, broadcast, merge, and arbitrary mapping for layers that enable construction of Neural Networks with complex topologies. The library tries really hard to avoid memory allocation and copying buffers, thus providing highly efficient and fast implementation.
This project is licensed under the MIT License - see the LICENSE file for details.