README.md

---
license: cc-by-4.0
library_name: pytorch
tags:
  - channel-modeling
  - ray-tracing
  - cir-prediction
  - vae
  - cvae
  - lightning
  - sionna
---

# CityMPC — TX_ViT pre-trained checkpoints (5 cities)

Stripped Lightning checkpoints for the **CityMPC TX_ViT** model from
the NeurIPS 2026 paper. One checkpoint per city, selected by the
seed-sweep stability filter (Kendall weighting σ_presence < 1e-3 +
good-verdict on test-set outlier rule), then ranked by lowest
Path Loss MAE / best Presence F1.

## Per-city metrics (test set)

| City | Path Loss MAE (dB) | ToF MAE (ns) | Avg Delay MAE (ns) | Presence F1 | Seed |
|---|---|---|---|---|---|
| tx_austin | 1.107 | 5.18 | 16.36 | 0.894 | 2024 |
| tx_dallas | 1.451 | 5.29 | 9.48 | 0.879 | 123 |
| tx_denver | 1.976 | 3.98 | 10.15 | 0.863 | 123 |
| tx_fortworth | 1.736 | 4.07 | 11.13 | 0.881 | 123 |
| tx_newyork | 2.481 | 8.59 | 19.96 | 0.874 | 42 |

## How to use

These checkpoints are stripped — **inference only**, not training-resumable.
After cloning the [CityMPC repo](https://github.com/neurips2026citympc/CityMPC):

```bash
# download a single city
hf download neurips2026citympc/citympc-txvit \
  outputs/tx_austin/city_10_austin_3p5_s/final/ckpts/last.ckpt \
  outputs/tx_austin/city_10_austin_3p5_s/final/config.yaml \
  --local-dir .

# evaluate
python scripts/analyze.py \
  --ckpt outputs/tx_austin/city_10_austin_3p5_s/final/ckpts/last.ckpt
```

Or open `notebooks/02_analyze_trained_model.ipynb` and paste the
downloaded run path into `TXVIT_RUN_DIR`.

## File layout

```
outputs/
  tx_austin/city_10_austin_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_dallas/city_8_dallas_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_denver/city_18_denver_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_fortworth/city_12_fortworth_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_newyork/city_0_newyork_3p5_s/final/{ckpts/last.ckpt, config.yaml}
```

## Integrity (SHA-256)

| File | Size | SHA-256 |
|---|---|---|
| `outputs/tx_austin/city_10_austin_3p5_s/final/ckpts/last.ckpt` | 60.6 MB | `2c4b31591eb347caab9acf11623dcc2d8a184515fbf8c4186b4958a2c5fe95a3` |
| `outputs/tx_austin/city_10_austin_3p5_s/final/config.yaml` | 0.0 MB | `f3e75a17d3d7726cfd01a75df635b26aac3ca54edd386af2b604eaaa7e3057c1` |
| `outputs/tx_dallas/city_8_dallas_3p5_s/final/ckpts/last.ckpt` | 60.6 MB | `397327907665bb7abb70335a0cd581677a8b6b55e9d3a4fbe95092d7fd9d12ad` |
| `outputs/tx_dallas/city_8_dallas_3p5_s/final/config.yaml` | 0.0 MB | `9b60b5483f2580737052ba8ed6e4d89fe86dd52161095b50a14afcbe7256a88d` |
| `outputs/tx_denver/city_18_denver_3p5_s/final/ckpts/last.ckpt` | 60.6 MB | `13e496219b7210f9092996023f16ef0aed53b066ac46b322cf8b664085706dee` |
| `outputs/tx_denver/city_18_denver_3p5_s/final/config.yaml` | 0.0 MB | `b03ea4570bbd423a719f91919293ed23617d48476476f7f398fb1ab0d3f51ea0` |
| `outputs/tx_fortworth/city_12_fortworth_3p5_s/final/ckpts/last.ckpt` | 60.6 MB | `6f6ce2e43b025edf422de5b94edc8c476dbc4d604137d487f00cf494c88f9317` |
| `outputs/tx_fortworth/city_12_fortworth_3p5_s/final/config.yaml` | 0.0 MB | `d7d124c9ff9d28c4ac8c7da2c3273863a4466e567480fae382593b5c9bd32e34` |
| `outputs/tx_newyork/city_0_newyork_3p5_s/final/ckpts/last.ckpt` | 60.6 MB | `98de5771b763445dd04d8c0bdf070777492710ffb0e65fa7807e599871b00e74` |
| `outputs/tx_newyork/city_0_newyork_3p5_s/final/config.yaml` | 0.0 MB | `c33baa28a3d6ae02626a46cd64f3161378cb53e4d93968528cc46091c861065d` |

## Notes

- Architecture: TX_ViT conditional VAE (CityMPC). See the paper for details.
- Trained on the [CityMPC dataset](https://huggingface.co/datasets/neurips2026citympc/CityMPC).
- Lightning state stripped (optimizer / scheduler / callback) — checkpoints are not training-resumable.
- σ-gate: every shipped run had `weighting/sigma/presence` < 1e-3 at its last logged step.