Aurora: Minimal Deep Learning Library.
Aurora is a minimal deep learning library written in Python, Cython, and C++ with the help of Numpy, CUDA, and cuDNN. Though it is simple, Aurora comes with some advanced design concepts found it a typical deep learning library.
- Automatic differentiation using static computational graphs.
- Shape and type inference.
- Static memory allocation for efficient training and inference.
Aurora relies on several external libraries including
cuDNN installation instructions please refer official documentation. Python dependencies can be installed by running the
To utilize GPU capabilities of the Aurora library, you need to have a Nvidia GPU. If
CUDA toolkit is not already installed, first install the latest version of the
CUDA toolkit as well as
cuDNN library. Next, set following environment variables.
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH export PATH=/usr/local/cuda/bin:$PATH
Cloning the Repository
You can clone Aurora repository using following command.
git clone https://github.com/upul/Aurora.git
Building the GPU Backend
Next, you need to build GPU backend. So please
cuda directory and run
make command as shown below.
- Go to
Installing the Library
Aurora directory and run:
pip install -r requirements.txt
pip install .
Following features will be added in upcoming releases.
- Dropout and Batch Normalization.
- High-level API similar to Keras.
- Ability to load pre-trained models.
- Model checkpointing.
It all started with CSE 599G1: Deep Learning System Design course. This course really helped me to understand fundamentals of Deep Learning System design. My answers to the two programming assignments of CSE 599G1 was the foundation of Aurora library. So I would like to acknowledge with much appreciation the instructors and teaching assistants of the SE 599G1 course.