Usage Guide

Learn how to use AZ-Go for training, playing games, and analyzing positions.

Basic Commands

Start Training 

# Main coordinator (on primary machine)
python start_main.py

# Worker nodes (on each worker)
python start_worker.py

# Note: Both scripts take no command-line arguments.
# Configuration is done through configs/config.yaml

Play Against the Model 

# Interactive GTP engine
python engine/run_engine.py       # Standard mode
python engine/run_engine.py -cli  # With command prompt

Analyze Games 

# Analyze SGF file with KataGo
python katago/run_katago.py  # Analyzes sgf/figure_c1.sgf (hardcoded)

Configuration Files

Main Configuration (configs/config.yaml) 

# Customize training parameters
board_size: 7

# Adjust neural network
nn_args:
  num_channels: 128
  depth: 18
  value_head_size: 32
  policy_head_size: 32

# Modify MCTS
num_mcts_sims: 500
cpuct: 1.0
fpu_value: 0.0

# Training settings
batch_size: 512
learning_rate: 0.01
epochs: 10

Note about Komi: The komi value is hardcoded in go/go_logic.py (line 25) as 5.5 and cannot be changed through the config file. To modify komi, you need to edit the source code directly.

Common Workflows

1. Training a New Model 

# Step 1: Configure parameters
nano configs/config.yaml

# Step 2: Start overseer (only 1 across all machines)
python start_main.py

# Step 3: Start workers (up to two per machine)
python start_worker.py

2. Resuming Training 

# Training automatically resumes from last checkpoint
# The overseer and workers handle checkpoint management internally
python start_main.py

Integration with Go GUIs

Sabaki

To use AZ-Go with Sabaki, you need to build a standalone executable:

  1. Configure the engine:
    • Place your model file as best.pth.tar in the engine/ directory
    • Edit engine/engine_config.yaml to set the correct network_type:
      • RES - 19 layer ResNet (current default)
      • DEP - Deprecated 18 layer ResNet (e.g., Model Q)
      • CNN - Convolutional Neural Network (not recommended)
  2. Build the GTP engine executable: 
    cd /path/to/AZ-Go
./engine/build_engine.sh

    This creates an executable in engine/dist/run_engine with the model embedded

  3. Add to Sabaki:
    • Open Sabaki → Engines → Manage Engines
    • Add new engine:
      • Path: /path/to/AZ-Go/engine/dist/run_engine
      • No arguments needed

Note: The engine at engine/run_engine.py is the GTP interface. The gtp/engine.py file is for analysis and visualization, not for playing.

Debugging and Analysis

Debug Scripts 

# Test self-play
python debug/debug_self_play.py

# Test arena
python debug/debug_arena.py

# Test scoring
python debug/debug_scoring.py

# Test worker self-play functionality
python debug/debug_worker.py