ShiroGo V2

Usage

1 Compile MXNet C++ Library

1.1 Requirements

• GCC or Clang
• Ninja
• CMake
• optional: CUDA
• The program has been tested on Linux and macOS.

1.2 Download MXNet

# download mxnet
mkdir -p ~/build/lib/
cd ~/build/lib/
wget https://github.com/apache/incubator-mxnet/releases/download/1.6.0/apache-mxnet-src-1.6.0-incubating.tar.gz
tar -xzvf apache-mxnet-src-1.6.0-incubating.tar.gz
mv apache-mxnet-src-1.6.0-incubating mxnet
cd mxnet

1.3.1 MacOS CPU

# cmake
cp CMakeLists_MacOS.txt CMakeLists.txt
mkdir build && cd build
cmake .. \
      -DCMAKE_BUILD_TYPE=Release \
      -DUSE_OPENMP=0 \
      -DUSE_CUDA=0 \
      -DUSE_OPENCV=0 \
      -DUSE_JEMALLOC=0 \
      -DUSE_CPP_PACKAGE=1 \
      -DUSE_BLAS=apple \
      -DUSE_F16C=0 \
      -GNinja

1.3.2 Linux CPU

# cmake
cp CMakeLists_Linux.txt CMakeLists.txt
mkdir build && cd build
cmake .. \
      -DCMAKE_BUILD_TYPE=Release \
      -DUSE_OPENMP=1 \
      -DUSE_CUDA=0 \
      -DUSE_OPENCV=0 \
      -DUSE_JEMALLOC=0 \
      -DUSE_CPP_PACKAGE=1 \
      -DUSE_F16C=0 \
      -GNinja

1.3.3 Linux GPU

# cmake
cp CMakeLists_Linux.txt CMakeLists.txt
mkdir build && cd build
cmake .. \
      -DCMAKE_BUILD_TYPE=Release \
      -DUSE_OPENMP=1 \
      -DUSE_CUDA=1 \
      -DUSE_CUDNN=1 \
      -DUSE_OPENCV=0 \
      -DUSE_JEMALLOC=0 \
      -DUSE_CPP_PACKAGE=1 \
      -DUSE_F16C=0 \
      -GNinja

1.4 Make

rm -rf lib && mkdir lib
cd build && ninja -v 
cp ./libmxnet* ../lib

1.5 Export mxnet

1.5.1 MacOS

export DYLD_LIBRARY_PATH=/Users/`yourname`/build/lib/mxnet/lib

1.5.2 Linux

export LD_LIBRARY_PATH=/home/`yourname`/build/lib/mxnet/lib

2 Compile ShiroGo

git clone https://github.com/yenw/ShiroGo.git
cd ShiroGo
mkdir -p build && cd build && cmake .. && make

3 Set Selfplay Task

cp -r tool/cr* ./build
cd build
# edit cr_auto
# ...
# ...

4 Run Selfplay

./cr_nohup_out

5 Watch Selfplay

./cr_watch

NN

Example

• NN Example block=2, filter=64.pdf

Supported Architecture

• Input
  • 16 channel: history planes
    • 8 step x 2 color
  • 2 channel: color planes
    • black
    • white
  • 2 channel: action planes
    • first: for normal move
    • second: for pass move
• Shared Body
  • Block
    • ResNet V1
    • ResNet V2
    • ResNext
    • ResNet used by CrazyZero
    • DPN: Dual Path Networks
    • MixNet
  • SE_Block
    • SE: squeeze & excitation
    • SNSE: SparseNet: squeeze & excitation
    • SA: Spatial Attention
    • CBAM: Channel attention & Spatial Attention
    • SAA: Spatial Attention & Add
    • KT
  • SE_Type:
    • Type A:
      • input2 = Conv3x3(input)
      • input3 = Conv3x3(input2)
      • output = SE(input3) + input
    • Type B:
      • input2 = Conv3x3(input)
      • input3 = Conv3x3(input2)
      • input4 = input3 + input
      • output = SE(input4) + input4
  • Activation
    • relu
    • elu
    • gelu
    • mish
    • swish1
    • leaky relu
    • prelu
    • rrelu
    • selu
• Output
  • Policy Head
    • Normal Move Head
    • Pass Head
  • Board Value Head
  • State Head
    • action = none, $state = s_t$
    • action = a, $state = s_{t+1} = P(s_t, a)$

N2N

• Support:
  • Shared Body Block: ResNetV2
  • Shared Body SE: none, KT-Type A, KT-Type B
  • Output: Policy, Board Value Head, State Head