Dragon: A Computation Graph Virtual Machine Based Deep Learning Framework

Compile Requirements for C++

1. Google Protocol Buffer
2. Python (2.7, 64bit)   |   Anaconda (2.7, 64bit)
3. CUDA [Optional]
4. CUDNN [Optional]
5. OpenMPI [Optional]

---

Runtime Requirements for Python

1. Package: protobuf
2. Package: lmdb

---

Installation

1. Clone this repository

2. (Optional) Download and install CUDA

   (Optional) Download and install CUDNN

3. (Optional) Download 3rdparty.zip and unzip to Dragon/3rdparty (Out of source code dir)

   Win64 (OpenBLAS / Protobuf for VS2013 / CUDNN v6 / Microsoft MPI)

   Linux64 (OpenMPI)

4. Configure Dragon/CMakeLists.txt

   • Select optional libraries [CUDA / CUDNN / BLAS / SSE / MPI / MPI_CUDA_AWARE / CUDA_FP16]
   • Set 3rdparty path (recommend to keep defualt)
   • Set python & numpy root path
   • Set cuda compiling architectures if necessary
   • GCC version(4.8+, 5.0-) should add -std=c++11 to CUDA_NVCC_FLAGS, if nullptr is not found.

5. Environment Variables

   Linux(Only for OpenMPI):

   • Create dragon.conf

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

     • Append 1 line for libraries dir of your 3rdparty, e.g. :
       • /home/Dragon/3rdparty/lib

   • rebuild the scaning cache

     sudo ldconfig

   Windows

   • add binary directionary to system environment variables, e.g. :
     • PATH=........;C:\Dragon\3rdparty\bin;

6. Setup 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

   • Run 3rdparty/setup_mpi.sh

     ./setup_mpi.sh

   • Install

     sudo cp openmpi/install/bin/mpirun /usr/bin

     Windows:

   • We use Microsoft MPI which can perfectly run at lastest Windows10

   • Microsoft MPI is intergrated into 3rdparty and you should do nothing

7. Compile

   Linux:

   • Install cmake

     sudo apt-get install cmake

   • Make

     cd Dragon
     mkdir build
     cd build
     cmake ..
     make install -j16

   Windows:

   • 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

8. Deploy

• Install Dragon

  cd Dragon
  python setup.py install

  Hint: If you do not have permission, try as follows:

  cd Dragon
  python setup.py install --user

• Install protobuf

  pip install protobuf

• Install lmdb

  pip install lmdb

---

Usage

Import

import dragon

Virtual DL Frameworks

import dragon.vm.theano as theano
import dragon.vm.caffe as caffe
import dragon.vm.tensorflow as tf

Tutorials

[IPython Notebook] -> (https://github.com/PhyscalX/Tutorials)

We will revise several classical examples, covering both CV, NLP and RL.

Device

import dragon.config
dragon.config.EnableCPU()
dragon.config.EnableCUDA(device_id, use_cudnn=True)

Automatic Memory Optimization(AMC)

import dragon.config
dragon.config.SetDebugMode(False)

This option will make all gradients share a global tensor(debugging is intractable).

which prefers a 50% memory-usage and 15% slower solution during training phase.

Scope

• NameScope

import dragon
from dragon.core.tensor import Tensor
with dragon.name_scope(prefix='conv1'):
    w = Tensor('weight').Variable()    # named as conv1/weight
    b = Tensor('bias').Variable()      # named as conv1/bias

• DeviceScope

import dragon
with dragon.deive_scope(deivce='gpu', id=0, use_cudnn=True):
    x = ops.Add(a, b)    # use /gpu:0 and cuDNN

• PhaseScope

import dragon
import dragon.vm.theano as theano
 with dragon.phase_scope(phase='train'):
    f = theano.function(outputs=y)    # force the training phase even without gradients computation

License and Citation

Dragon is released under the BSD 2-Clause license.

Please cite Dragon in your publications if it helps your research:

@article{pan2017dragon,
  Author = {Pan, Ting},
  Journal = {arXiv preprint arXiv:1707.08265},
  Title = {Dragon: A Computation Graph Virtual Machine Based Deep Learning Framework},
  Year = {2017}
}