(or: 500 ways to train deep neural networks)

Deep500 is a library that can be used to customize and measure anything with deep neural networks, using a clean, high-performant, and simple interface. Deep500 includes four levels of abstraction: (L0) Operators (layers); (L1) Network Evaluation; (L2) Training; and (L3) Distributed Training.

Using Deep500, you automatically gain:

• Operator validation, including gradient checking for backpropagation
• Statistically-accurate performance benchmarks and plots
• High-performance integration with popular deep learning frameworks (see Supported Frameworks below)
• Running your operator/framework/optimizer/communicator/... with real workloads, alongside existing environments
• and much more...

Using pip: pip install deep500

See the tutorials.

• Python 3.5 or later
• Protobuf (sudo apt-get install protobuf-compiler libprotoc-dev)
• For plotted metrics: matplotlib
• For distributed optimization:
  • Any MPI implementation (OpenMPI, MPICH, MVAPICH etc.)
  • mpi4py Python package

• Tensorflow
• Pytorch
• Caffe2

If you use this meta-framework please cite it as:

@inproceedings{deep500,
  author={T. Ben-Nun and M. Besta and S. Huber and A. N. Ziogas and D. Peter and T. Hoefler},
  title={{A Modular Benchmarking Infrastructure for High-Performance and Reproducible Deep Learning}},
  year={2019},
  month={May},
  publisher={IEEE},
  note={The 33rd IEEE International Parallel \& Distributed Processing Symposium (IPDPS'19)},
}

Deep500 is an open-source, community driven project. We are happy to accept Pull Requests with your contributions!

Deep500 is published under the New BSD license, see LICENSE.