August practice: 16 CPU-only generative-model projects on synthetic calorimeter showers — flow matching, VAE-style latents, evaluation benchmarks.
This repository contains 16 standalone CPU-only projects (p5–p20) exploring generative modeling on synthetic calorimeter shower data. All projects run on NumPy / PyTorch / matplotlib and are validated through GitHub Actions CI on every push.
| # | Project | Highlight |
|---|---|---|
| p5 | calo-demo | NumPy calorimeter shower simulator: profiles, observables, event displays |
| p6 | conditional-cfm | Conditional flow matching on beam energy, C2ST AUC 0.998 |
| p7 | energy-interpolation | Sample unseen beam energies (30/70 GeV) from a flow trained on 10/50/100 GeV |
| p8 | flow-distillation | Distilled 50-step flow sampler into a 1-step student, 45× faster |
| p9 | pointcloud-cfm | Point-cloud shower generation with a DeepSets velocity net (41k params) |
| p10 | c2st-test | Classifier two-sample test: physics features vs raw pixels |
| p11 | ensemble-flow | 3-member flow ensemble with epistemic uncertainty maps |
| p12 | attn-unet-flow | Attention U-Net vs ResMLP velocity field comparison |
| p13 | layer-autoregressive | Layer-by-layer autoregressive shower generation (4 stages) |
| p14 | quantized-flow | INT8 quantisation + pruning: 3.8× smaller with ~zero AUC loss |
| p15 | benchmark-3-gens | WGAN-GP vs DDPM vs CFM on identical data |
| p16 | anomaly-flow | Flow log-prob as anomaly score (negative result, AUC 0.486) |
| p17 | latent-cfm | VAE latent space + flow matching in latent space |
| p18 | cfg-calo | Classifier-free guidance for energy control (w sweep 0-7) |
| p19 | transformer-cfm | DiT-style transformer velocity field (AdaLN) |
| p20 | consistency-calo | Consistency model, 1-step sampling |
git clone https://github.com/AlphaCortex-03/Gen-Ai--Aug26.git
cd Gen-Ai--Aug26
pip install -r requirements.txt
pip install torch --index-url https://download.pytorch.org/whl/cpuRun a single project:
cd p6-conditional-cfm
python src/main.py # or src/train.py depending on projectRun all projects sequentially (smoke mode, fast):
python run_all.py --smokeSee python run_all.py --help for --only, --skip, --stop, --list, and --timeout options.
GitHub Actions smoke-runs all 16 projects on every push and pull request. Results are uploaded as run_results.json.
See CONTRIBUTING.md for how to report bugs, suggest features, and submit pull requests.
See CODE_OF_CONDUCT.md.
This project is licensed under the MIT License — see LICENSE.