---
license: mit
language:
  - en
pipeline_tag: reinforcement-learning
tags:
  - counter-strike-2
  - trajectory-prediction
  - transformer
  - esports
---

# CS-NET v4 — CS2 Path Prediction & Spatial-Only Models

Trained checkpoints for **CS-NET v4** — a Transformer-based framework for
Counter-Strike 2 game-state forecasting. Full open-source code (training,
inference, 3D visualizer) is available at
**[github.com/Gary2005/cs-net](https://github.com/Gary2005/cs-net)**.

All checkpoints use the **Pro architecture** (d_model=768, ~138.7M params),
see `config/pretrain-a100-pro.yaml` in the source repo.

## Files

| File | Task | Description |
|---|---|---|
| `cs-net-v4-pro.pt` | Path prediction (pre-training) | 600k steps. Full model: `{model, global_step}`. Auto-regressively predicts each player's future movement path (16 ticks = 4 s) from a window of the full game state (10 players + bomb + projectiles + raycast depth maps). |
| `pretrain-v4-pro-win_rate.pt` | Spatial-only `winrate` | Per-player team win probability from a **single tick** of the game state. `{task, model_state, head_state, config}` — embedder + spatial transformer + linear head. |
| `pretrain-v4-pro-alive_end.pt` | Spatial-only `alive_end` | Per-player probability of being alive at round end. Same format. |
| `pretrain-v4-pro-future_kill.pt` | Spatial-only `future_kill` | Per-player probability of getting a kill after the current tick. Same format. |

## Quick start

```bash
git clone https://github.com/Gary2005/cs-net.git
cd cs-net
conda create -n cs2demo python=3.10 && conda activate cs2demo
pip install -r requirements.txt
pip install huggingface_hub

# download all 4 checkpoints
python scripts/download_checkpoints.py --repo-id gary2oos/cs-net-v4

# verify they load correctly (architecture match, finite weights, forward pass)
python scripts/test_checkpoints.py --models-dir checkpoints
```

### Inference

```python
# Path prediction
from scripts.prediction_engine import PredictionEngine
engine = PredictionEngine("config/pretrain-a100-pro.yaml",
                          "checkpoints/cs-net-v4-pro.pt",
                          device="cpu", maps_dir="maps/optimized_obj_files")
result = engine.predict_at_tick(sample, query_tick=120)  # sample = round-level dict

# Spatial-only (winrate / alive_end / future_kill)
from scripts.spatial_only_predictor import SpatialOnlyPredictor
predictor = SpatialOnlyPredictor("checkpoints", device="cpu")
out = predictor.predict_round_full(sample)   # per-tick per-player probabilities
```

### 3D visualizer

```bash
python visualizer/server.py --port 5000 \
    --checkpoint checkpoints/cs-net-v4-pro.pt \
    --spatial-model-dir checkpoints \
    --device cpu
# open http://127.0.0.1:5000/
```

## License

MIT (see source repo).