argman / Dragon

Dragon: A Light Deep Learning Framework (0.1rc)

Dragon: A Light Deep Learning Framework

Specifics !

• Light 3rdparty: Protobuf / Boost and Install Within 10 Minutes

• CaffeModels Support

• Distributed Architectures Support

  • Device Parallelism: Use muti-GPUs to run a huge Neural Network(e.g. Res-Net)
  • Data Parallelism: Update several Neural Networks with a ParameterServer
  • Python Parallelism: Be compatible with Python

• Symbols Support (in development)

• Cross Platforms (Linux / Windows) and United Deployment

How to Install ?

Requirements

1. CUDA
2. Anaconda [Optional]
3. Microsoft MPI / OpenMPI [Optional]

Installation

1. clone this repository

2. download and install CUDA

   • we recommend install CUDA8 or higher to support GCC5/VS2015

3. download 3rdparty.zip and unzip to any path

   • Win64_VS2013_Release
   • Win64_VS2015_Release
   • Linux64_GCC4 (Use GCC5 but compile libprotobuf with GCC4)
   • Linux64_GCC5
   • we recommond unzip to /home/xxx/3rdparty (Linux) or C:/3rdparty (Windows)

4. install Anaconda with x64-py2.7 [Optional]

   • allow importing environment variable into bashrc (Linux) / PATH (Windows)
   • install protobuf

   pip install protobuf

5. configure Dragon/CMakeLists.txt

   • select optional libraries [PYTHON / MPI / CUDA_AWARE_MPI]
   • set 3rdparty path
   • set python path [Optional]
   • set cuda compiling architectures if necessary

6. set environment variables

   • Linux:
     • create dragon.conf

      sudo gedit /etc/ld.so.conf.d/dragon.conf

     • append 2 lines with ld dirs for your 3rdparty, e.g:
       • /home/xxx/3rdparty/bin
       • /home/xxx/3rdparty/lib
     • rebuild the scaning cache

      sudo ldconfig

   • Windows
     • add 3rdparty/bin to system environment variables, e.g:
       • PATH=........;C:\3rdparty\bin;

7. install MPI [Optional]

   • Linux:
     • we use OpenMPI which support "cuda-aware-mpi"
     • see more:
       • https://devblogs.nvidia.com/parallelforall/introduction-cuda-aware-mpi/
       • https://www.open-mpi.org/faq/?category=buildcuda
     • configure 3rdparty/setup_mpi.sh
       • disable cuda-aware support if necessary
     • run 3rdparty/setup_mpi.sh

      sudo ./setup_mpi.sh

   • Windows:
     • we use Microsoft MPI which can perfectly run at lastest Windows10
     • Microsoft MPI is intergrated into 3rdparty and you should do nothing

8. compile

   • Linux:
     • install cmake

      sudo apt-get install cmake

     • configure Dragon/main_install.sh
       • set 3rdparty path
     • run Dragon/main_install.sh

      sudo ./main_install.sh

   • Windows:
     • run Dragon/bin/gen_proto.bat
     • install cmake-gui
     • mkdir Dragon/build
     • configure and generate MSVC project in Dragon/build
     • open Dragon/build/Dragon.sln
     • compile and generate for "INSTALL" solution

9. install PyDragon [Optional]

   • Linux:
     • configure Dragon/py_install.sh
       • set 3rdparty path
       • set anaconda path
     • run Dragon/py_install.sh

      sudo ./py_install.sh

   • Windows:
     • configure Dragon/py_install.bat
       • set 3rdparty path
       • set anaconda path
     • run Dragon/py_install.bat

Hints

Developer: PanTing, HeFei University of Technology at Xuancheng

We will change the codes frequently

This repo is also regarded as the Developer's Bachelor Graduation Project

License and Citation

Dragon is a distribution of the Caffe(BVLC)、MPI-Caffe(Stefan Lee)

Please cite their projects firstly in your publications if it helps your research