CS-NET

A deep learning framework for Counter-Strike match data analysis

中文文档

Quick Links

• Project Overview
• Prediction Tasks
• Quick Start
• Web App Usage
• Web App Features
• Acknowledgements
• Contributors

---

📌 Project Overview

CS-NET is a Transformer-based deep learning framework for analyzing Counter-Strike 2 match replays (.dem demo files). It parses match recordings, converts game states into token sequences, and uses pre-trained Transformer models for multiple real-time predictions.

In short: given a match replay, the model tells you who will win, who will die, and who is most likely to get the next kill.

Prediction Tasks

Task	Description	Output
Win Rate Prediction	Probability of team1 (mapped from CT/T by side) winning the current round	Scalar in [0, 1]
Alive Prediction	Per-player probability of surviving the next 5 seconds	One probability per player
Next Kill Prediction	Which player is most likely to get the next kill	Probability distribution over 10+1 classes
Next Death Prediction	Which player is most likely to die next	Probability distribution over 10+1 classes
Duel Prediction	1v1 win probability for any CT-T player pair	5x5 probability matrix

🚀 Quick Start

1. Setup Environment

Create a Python environment and install dependencies:

conda create -n cs-net python=3.10
conda activate cs-net
pip install -r requirements.txt

2. Download Pre-trained Models

Download all pre-trained models and tokenizers to ./cs-net-models/:

Model weights are also available here: https://huggingface.co/gary2oos/CS-Net-V3

python -m scripts.download_model

3. Convert Demo to JSON

Process a Counter-Strike demo file (.dem) into structured JSON format:

examples/test.dem is intentionally NOT included in this repository because demo files are too large. You must download a .dem file yourself (for example from HLTV) and replace the input path.

python -m data.process_demo \
  -path examples/test.dem \
  -interval 0.25 \
  -out examples/test.json

4. Download Test Data

To reproduce the calibration/evaluation numbers below, download the test shard first:

python -m scripts.download_data

This script downloads test/shards-00000.tar from Hugging Face and stores it under ./dataset/test/.

5. Calibrate Temperature Scaling

You can calibrate the model3.0 heads with:

python -m scripts.train3_t_scaling --dataset_path dataset --device cpu

On the current test shard, the calibration results are:

Task	T	Before Loss	Before ECE	Before Acc	After Loss	After ECE	After Acc
Alive	1.193158	0.431486	0.028640	0.774641	0.429344	0.021371	0.774641
Duel	1.146975	0.633122	0.017835	0.632217	0.632133	0.014517	0.632217
Next Death	1.493995	1.785793	0.077382	0.342485	1.741089	0.012107	0.342485
Next Kill	1.602551	1.801647	0.102502	0.339024	1.736153	0.012922	0.339024
Win Rate	1.061342	0.467820	0.029351	0.754566	0.467459	0.029516	0.754566

🌐 Web App Usage

CS-NET now includes an interactive web app for demo analysis and LLM-based post-game summary.

Attribution Notice The built-in 2D viewer is a modified integration of sparkoo/csgo-2d-demo-viewer. We use the upstream project under the MIT License and adapt it for CS-NET's Flask routes and model-prediction overlays.

1. Start the web app

python -m demo_analysis.web_app

Then open:

http://127.0.0.1:7860

2. Analyze a demo in UI

1. Upload a .dem file.
2. Select the model root directory (normally cs-net-models/). The web app loads all five prediction heads (alive, nxt_kill, nxt_death, win_rate, duel) from their subdirectories in one go.
3. Select device (cpu / cuda).
4. Click Start Analysis.

3. Generate LLM summary

1. Fill API Key, model name, and Base URL (OpenAI-compatible).
2. Choose app language (Chinese / English).
3. Click Generate AI Review.

✨ Web App Features

• Bilingual UI and bilingual LLM output (Chinese / English).
• Round-by-round win-rate curve with kill markers.
• Hover-to-inspect player contribution at each timeline point.
• Live 2D radar that syncs with the win-rate curve — player positions, team colour, alive/dead state, and "recently flashed" flag are drawn on the real minimap overview for every tick the cursor touches.
• Per-tick metric panels driven by all four prediction heads: 5-second survival probability, next-kill distribution, next-death distribution, and the full 5×5 CT-vs-T duel matrix.
• Advanced metrics table aggregated across the whole match: per-player average kill/death/survival probability, hard-duel win rate (fights the model thought they would lose), easy-duel win rate (fights they were favoured in), highlight rate, plus a |swing|-sorted ranking of the most impactful kills.
• One-click 2D replay viewer — launches a bundled build of sparkoo/csgo-2d-demo-viewer in a new tab with the same demo, adding smoke/flash/grenade trajectories and an in-page timeline overlaid with CS-NET's predictions.
• Current round final contribution table + full match average contribution table.
• MVP and SVP badges.
• LLM summary supports streaming output and Markdown rendering.
• Auto-save user settings in browser local storage: API Key, model name, base URL, temperature, device, model path, batch size, language.
• Team-side context for LLM: per-round attack/defense roles, first-half/second-half side assignment and half scores.

🙏 Acknowledgements

The bundled 2D replay viewer under demo_analysis/static/viewer/ is a lightly modified build of the excellent open-source project sparkoo/csgo-2d-demo-viewer by Michal Vala, distributed under the MIT License (© 2023 Michal Vala). All credit for the viewer's parsing, rendering, and UX belongs to the upstream authors — CS-NET only rewires its asset paths and feeds in the per-tick predictions from our models. Huge thanks to Michal and the upstream contributors for making such a polished tool available to the community.

The original upstream license is reproduced verbatim at demo_analysis/static/viewer/LICENSE and applies to every file in that directory. If you reuse or redistribute the viewer portion of this repository, please preserve that notice.

⭐️ Star History

🤝 Contributors

• Gary2005
• czdzx
• Yianlaen
• likedeedlit-eng