Deep500: 深度学习元框架和HPC基准测试库

Deep500: 深度学习元框架和HPC基准测试库

Deep500: A Deep Learning Meta-Framework and HPC Benchmarking Library

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 backpropagationStatistically-accurate performance benchmarks and plotsHigh-performance integration with popular deep learning frameworks (see Supported Frameworks below)Running your operator/framework/optimizer/communicator/... with real workloads, alongside existing environmentsand much more...

Installation

Using pip: pip install deep500

Usage

See the tutorials.

Requirements

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

Supported Frameworks

TensorflowPytorchCaffe2

Reference

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)},}

Contributing

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

License

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

版权声明：本文内容由网络用户投稿，版权归原作者所有，本站不拥有其著作权，亦不承担相应法律责任。如果您发现本站中有涉嫌抄袭或描述失实的内容，请联系我们jiasou666@gmail.com 处理，核实后本网站将在24小时内删除侵权内容。