Go (바둑) model for 9x9 board
96 filters of 3x3 kernels, 18 residual blocks.
Policy Head Features (2x9x9) - Plane 1
Policy Head Features (2x9x9) - Plane 2
Network Interpretability
To see what the model is focusing on, I visualized the activations from the policy head's 2x9x9 intermediate layer. This is the final abstract representation of the board state right before the network flattens the data to calculate move probabilities.
Above are the live internal features for the current board state, visualized in three different ways: Heatmap (raw activations), Influence Glow (smoothed attention), and Hint Circle (peak activation points).
The model was trained on high-rank games from the GoQuest server using a publicly available dataset. Training was performed with mixed precision (bfloat16) and detailed logs are available in the WandB Report (Run: 20251218_182225).
Architecture & Optimization
- Structure: 18 residual blocks with 96 filters (3x3 kernels).
- Policy Head: No search or rollouts are performed. The policy head transforms the board into 2x9x9 features, followed by a fully connected layer outputting 82 probabilities (81 moves + pass).
- Deployment: The model was exported to ONNX and compressed using dynamic quantization (unsigned 8-bit integers).
Performance Impact
Quantization made the model way lighter to deploy both in terms of size and speed, and didn't really cost any accuracy:
- Model Size: Reduced from 12.1 MB → 3.2 MB.
- Val Set Move Prediction Top-1 Accuracy: 60.42% → 60.28%.
- Val Set Move Prediction Top-3 Accuracy: 86.71% → 86.70%.