KataGo Training Mechanism

This article provides an in-depth analysis of KataGo's training mechanism, helping you understand how self-play training works.

---

Training Overview

Training Loop

Initial model → Self-play → Collect data → Train update → Stronger model → Repeat

Animation correspondence:

• E5 Self-play ↔ Fixed point convergence
• E6 Strength curve ↔ S-curve growth
• H1 MDP ↔ Markov chain

Hardware Requirements

Model Scale	GPU Memory	Training Time
b6c96	4 GB	Several hours
b10c128	8 GB	1-2 days
b18c384	16 GB	1-2 weeks
b40c256	24 GB+	Several weeks

---

Environment Setup

Install Dependencies

# Python environment
conda create -n katago python=3.10
conda activate katago

# PyTorch (CUDA version)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

# Other dependencies
pip install numpy h5py tqdm tensorboard

Get Training Code

git clone https://github.com/lightvector/KataGo.git
cd KataGo/python

---

Training Configuration

Config File Structure

# configs/train_config.yaml

# Model architecture
model:
  num_blocks: 10          # Number of residual blocks
  trunk_channels: 128     # Trunk channel count
  policy_channels: 32     # Policy head channels
  value_channels: 32      # Value head channels

# Training parameters
training:
  batch_size: 256
  learning_rate: 0.001
  lr_schedule: "cosine"
  weight_decay: 0.0001
  epochs: 100

# Self-play parameters
selfplay:
  num_games_per_iteration: 1000
  max_visits: 600
  temperature: 1.0
  temperature_drop_move: 20

# Data settings
data:
  max_history_games: 500000
  shuffle_buffer_size: 100000

Model Scale Reference

Name	num_blocks	trunk_channels	Parameters
b6c96	6	96	~1M
b10c128	10	128	~3M
b18c384	18	384	~20M
b40c256	40	256	~45M

Animation correspondence:

• F2 Network size vs strength: Capacity scaling
• F6 Neural scaling laws: Log-log relationship

---

Training Process

Step 1: Initialize Model

# init_model.py
import torch
from model import KataGoModel

config = {
    'num_blocks': 10,
    'trunk_channels': 128,
    'input_features': 22,
    'policy_size': 362,  # 361 + pass
}

model = KataGoModel(config)
torch.save(model.state_dict(), 'model_init.pt')
print(f"Model parameters: {sum(p.numel() for p in model.parameters()):,}")

Step 2: Self-Play Data Generation

# Compile C++ engine
cd ../cpp
mkdir build && cd build
cmake .. -DUSE_BACKEND=CUDA
make -j$(nproc)

# Run self-play
./katago selfplay \
  -model ../python/model_init.pt \
  -output-dir ../python/selfplay_data \
  -config selfplay.cfg \
  -num-games 1000

Self-play configuration (selfplay.cfg):

maxVisits = 600
numSearchThreads = 4

# Temperature settings (increase exploration)
chosenMoveTemperature = 1.0
chosenMoveTemperatureEarly = 1.0
chosenMoveTemperatureHalflife = 20

# Dirichlet noise (increase diversity)
rootNoiseEnabled = true
rootDirichletNoiseTotalConcentration = 10.83
rootDirichletNoiseWeight = 0.25

Animation correspondence:

• C3 Exploration vs exploitation: Temperature parameter
• E10 Dirichlet noise: Root exploration

Step 3: Train Neural Network

# train.py
import torch
from torch.utils.data import DataLoader
from model import KataGoModel
from dataset import SelfPlayDataset

# Load data
dataset = SelfPlayDataset('selfplay_data/')
dataloader = DataLoader(dataset, batch_size=256, shuffle=True)

# Load model
model = KataGoModel(config)
model.load_state_dict(torch.load('model_init.pt'))
model = model.cuda()

# Optimizer
optimizer = torch.optim.Adam(
    model.parameters(),
    lr=0.001,
    weight_decay=0.0001
)

# Learning rate scheduler
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
    optimizer,
    T_max=100,
    eta_min=0.00001
)

# Training loop
for epoch in range(100):
    model.train()
    total_loss = 0

    for batch in dataloader:
        inputs = batch['inputs'].cuda()
        policy_target = batch['policy'].cuda()
        value_target = batch['value'].cuda()
        ownership_target = batch['ownership'].cuda()

        # Forward pass
        policy_pred, value_pred, ownership_pred = model(inputs)

        # Compute loss
        policy_loss = torch.nn.functional.cross_entropy(
            policy_pred, policy_target
        )
        value_loss = torch.nn.functional.mse_loss(
            value_pred, value_target
        )
        ownership_loss = torch.nn.functional.mse_loss(
            ownership_pred, ownership_target
        )

        loss = policy_loss + value_loss + 0.5 * ownership_loss

        # Backward pass
        optimizer.zero_grad()
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()

        total_loss += loss.item()

    scheduler.step()
    print(f"Epoch {epoch}: Loss = {total_loss / len(dataloader):.4f}")

    # Save checkpoint
    torch.save(model.state_dict(), f'model_epoch{epoch}.pt')

Animation correspondence:

• D5 Gradient descent: optimizer.step()
• K2 Momentum: Adam optimizer
• K4 Learning rate decay: CosineAnnealingLR
• K5 Gradient clipping: clip_grad_norm_

Step 4: Evaluation & Iteration

# Evaluate new model vs old model
./katago match \
  -model1 model_epoch99.pt \
  -model2 model_init.pt \
  -num-games 100 \
  -output match_results.txt

If new model win rate > 55%, replace old model and proceed to next iteration.

---

Loss Functions Explained

Policy Loss

# Cross-entropy loss
policy_loss = -sum(target * log(pred))

Goal: Make predicted probability distribution close to MCTS search results.

Animation correspondence:

• J1 Policy entropy: Cross-entropy
• J2 KL divergence: Distribution distance

Value Loss

# Mean squared error
value_loss = (pred - actual_result)^2

Goal: Predict final game result (win/loss/draw).

Ownership Loss

# Per-point ownership prediction
ownership_loss = mean((pred - actual_ownership)^2)

Goal: Predict final ownership of each position.

---

Advanced Techniques

1. Data Augmentation

Leverage board symmetry:

def augment_data(board, policy, ownership):
    """Data augmentation using D4 group's 8 transformations"""
    augmented = []

    for rotation in range(4):
        for flip in [False, True]:
            # Rotation and flip
            aug_board = transform(board, rotation, flip)
            aug_policy = transform(policy, rotation, flip)
            aug_ownership = transform(ownership, rotation, flip)
            augmented.append((aug_board, aug_policy, aug_ownership))

    return augmented

Animation correspondence:

• A9 Board symmetry: D4 group
• L4 Data augmentation: Symmetry exploitation

2. Curriculum Learning

From simple to complex:

# Train with fewer search visits first
schedule = [
    (100, 10000),   # 100 visits, 10000 games
    (200, 20000),   # 200 visits, 20000 games
    (400, 50000),   # 400 visits, 50000 games
    (600, 100000),  # 600 visits, 100000 games
]

Animation correspondence:

• E12 Training curriculum: Curriculum learning

3. Mixed Precision Training

from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler()

with autocast():
    policy_pred, value_pred, ownership_pred = model(inputs)
    loss = compute_loss(...)

scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()

4. Multi-GPU Training

import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel

# Initialize distributed
dist.init_process_group(backend='nccl')

# Wrap model
model = DistributedDataParallel(model)

---

Monitoring & Debugging

TensorBoard Monitoring

from torch.utils.tensorboard import SummaryWriter

writer = SummaryWriter('runs/training')

# Log losses
writer.add_scalar('Loss/policy', policy_loss, step)
writer.add_scalar('Loss/value', value_loss, step)
writer.add_scalar('Loss/total', total_loss, step)

# Log learning rate
writer.add_scalar('LR', scheduler.get_last_lr()[0], step)

tensorboard --logdir runs

Common Issues

Issue	Possible Cause	Solution
Loss not decreasing	Learning rate too low/high	Adjust learning rate
Loss oscillating	Batch size too small	Increase batch size
Overfitting	Insufficient data	Generate more self-play data
Strength not improving	Too few search visits	Increase maxVisits

Animation correspondence:

• L1 Overfitting: Over-adaptation
• L2 Regularization: weight_decay
• D6 Learning rate effects: Tuning

---

Small-Scale Experiment Suggestions

If you just want to experiment, we recommend:

1. Use 9×9 board: Dramatically reduces computation
2. Use small model: b6c96 is enough for experiments
3. Reduce search visits: 100-200 visits
4. Fine-tune pretrained model: Faster than training from scratch

# 9×9 board settings
boardSize = 9
maxVisits = 100

---

Further Reading

• Source Code Guide — Understanding code structure
• Contributing to Open Source — Join distributed training
• KataGo's Key Innovations — The secret to 50x efficiency