Files
sor/tools/README.md
T
d5a0efdb77 Implement AI Phase 5: Python neural training pipeline
- tools/train_neural.py: PyTorch MLP training script that reads JSONL from
  GameReplayExporter and exports a compliant ONNX model. Architecture is
  1024->512->256->128->1 (tanh) with LayerNorm + Dropout. AdamW optimizer
  with cosine LR annealing, gradient clipping, best-checkpoint restore.
  Post-export validation via onnxruntime confirms tensor names and shapes.
- tools/requirements.txt: pinned deps (torch 2.3.1, onnx 1.16.1,
  onnxruntime 1.18.0 matching the C# package, numpy 1.26.4)
- tools/README.md: end-to-end workflow docs covering self-play, data export,
  training options, deployment (hot-swap vs restart), ONNX contract table,
  and iterative training guidance

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
2026-04-12 21:26:45 -04:00

146 lines
4.1 KiB
Markdown

# SOR Neural AI Training Tools
This directory contains the Python training pipeline for the SOR neural decision engine
(tiers 9-10). The workflow is:
```
self-play (admin) -> JSONL export (admin) -> train_neural.py -> model.onnx -> hot-swap (admin)
```
---
## Prerequisites
Python 3.11+ with pip:
```bash
pip install -r tools/requirements.txt
```
GPU training is supported via CUDA. Use `--cpu` to force CPU-only mode.
---
## Step 1 — Generate self-play data (Admin portal)
1. Navigate to **Admin -> AI -> Calibration**.
2. Click **Run All Pairs** (or a specific difficulty pair) to trigger in-process self-play games.
3. Wait for the run to complete — check the calibration table for results.
Alternatively, trigger a full calibration run via the API:
```
POST /api/ai/calibration/run-all
```
---
## Step 2 — Export training data (Admin portal)
1. Navigate to **Admin -> AI -> Calibration**.
2. Click **Export Training Data** and enter an output path on the server
(e.g. `/var/sor/training/data.jsonl`).
3. The exporter replays completed game sessions and writes one JSON line per action:
```json
{
"GameSessionId": "3fa85f64...",
"TurnNumber": 5,
"ActorAccountId": "...",
"Features": [0.0, 1.0, ...],
"ActionType": "AiPlayCardAction",
"ActionJson": "{}",
"Outcome": 1.0
}
```
`Outcome` is `+1.0` if the acting player won the game, `-1.0` if they lost.
---
## Step 3 — Train the model
```bash
cd /path/to/SOR
python tools/train_neural.py --data /var/sor/training/data.jsonl --out /var/sor/models/model.onnx --epochs 50 --batch-size 512 --lr 1e-3
```
### Options
| Flag | Default | Description |
|------|---------|-------------|
| `--data` | *(required)* | Path to the JSONL export file |
| `--out` | `model.onnx` | Output ONNX file path |
| `--epochs` | `50` | Number of training epochs |
| `--batch-size` | `512` | Mini-batch size |
| `--lr` | `0.001` | AdamW initial learning rate |
| `--val-split` | `0.1` | Fraction of data held out for validation |
| `--dropout` | `0.2` | Dropout probability in the MLP |
| `--seed` | `42` | Random seed |
| `--cpu` | off | Force CPU training (ignores CUDA) |
The script prints per-epoch train/val loss and marks the best checkpoint with `*`.
It restores the best weights before exporting.
### Model architecture
```
Input [1, 1024] float32 (FeatureExtractor output)
Linear 1024 -> 512 + LayerNorm + ReLU + Dropout
Linear 512 -> 256 + LayerNorm + ReLU + Dropout
Linear 256 -> 128 + LayerNorm + ReLU
Linear 128 -> 1 + Tanh
Output [1, 1] float32 (score in [-1, 1])
```
Total parameters: ~920k
---
## Step 4 — Deploy the model
### Option A: Hot-swap via admin portal (no restart)
1. Copy the `model.onnx` file to the path configured in `appsettings`:
```json
"Ai": { "NeuralModelPath": "/var/sor/models/model.onnx" }
```
2. Navigate to **Admin -> AI -> Neural Model**.
3. Click **Hot-swap Model**. The old session is disposed atomically and the new
model is loaded from disk. All subsequent tier 9-10 decisions use the new model.
### Option B: Restart the service
The model is loaded at startup by `AiNeuralStartupService`. If the file exists at the
configured path, it will be loaded automatically on the next deploy.
---
## ONNX contract
The exported model must satisfy:
| Property | Value |
|----------|-------|
| Input tensor name | `features` |
| Input shape | `[1, 1024]` (or `[batch, 1024]` with dynamic batch) |
| Input dtype | `float32` |
| Output tensor name | `score` |
| Output shape | `[1, 1]` (or `[batch, 1]`) |
| Output dtype | `float32` |
| Output range | `[-1, 1]` (tanh) |
These names are defined as constants in `NeuralModelRegistry.cs` (`InputName`, `OutputName`).
The Python script and the C# inference code must agree on these names exactly.
---
## Iterative training
For best results, run multiple rounds:
1. Start with random-engine games (all difficulties) to generate diverse data.
2. Train an initial model and deploy it.
3. Run additional self-play with the neural engine enabled (tiers 9-10 play each other).
4. Append new exports to the existing JSONL and retrain with more epochs.
5. Monitor calibration win rates in the admin portal after each deployment.