# Scenario-Aware Multi-Task

**Repository Path**: bistu_liuning_admin_admin/scenario-aware-multi-task

## Basic Information

- **Project Name**: Scenario-Aware Multi-Task
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: GPL-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-07-19
- **Last Updated**: 2026-07-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# NK-Fallback: code repository

[![Python 3.10+](https://img.shields.io/badge/Python-3.10%2B-3776AB)](https://www.python.org/)
[![PyTorch 2.1+](https://img.shields.io/badge/PyTorch-2.1%2B-EE4C2C)](https://pytorch.org/)

This repository contains the complete data-processing, training,
evaluation and figure-generation pipeline for a scenario-aware
multi-task learning framework on real-world autonomous-vehicle
fallback data.  It accompanies a separate IEEE T-IV submission
(not redistributed in this public repo).

## 1. Repository layout

```
beike-paper/
+- 01_*.py .. 34_*.py       Numbered scripts (see docs/PIPELINE.md for the
|                           end-to-end data + train + figure pipeline)
+- 05_npz/                  Per-segment decoded npz (gitignored, regenerable)
+- 06_features/             features.npz (state, action, future targets)
|                           (gitignored, regenerable)
+- 07_results/              Per-experiment JSON metrics
+- docs/
   +- PIPELINE.md           Script-by-script map of the full pipeline
   +- HARDWARE.md           Reproduce on the BISTU RTX-4090 server
+- README.md                this file
+- requirements.txt
+- .gitignore
+- run_pipeline.sh / .bat   one-shot end-to-end driver
```

## 2. Quick start

```bash
# 1. install deps (CUDA 12.x assumed)
pip install -r requirements.txt

# 2. re-create features.npz from raw bags/CSV in S:\01_\ ...
python 05_decode_all.py        # ~12 s per bag, ~3 min for 27 segments
python 06_features.py           # ~3 s, writes 06_features/features.npz

# 3. train all 12 evaluation runs (random / cross-date / day-night / LOSO)
python 23_run_experiments.py    # uses local GPU; 22 s per run

# 4. train the 6 ablation runs (backbone / scene conditioning)
python 27_run_ablation.py

# 5. regenerate every figure (English labels) for offline review
python 34_regen_all_figures_en.py
```

End-to-end wall-clock on a single RTX 4090 is **under 5 minutes** (data
re-decode excluded).

## 3. Reproducing on the BISTU GPU server (10.1.58.237)

`docs/HARDWARE.md` documents the SSH / SFTP path used in this work
(scripts 13-22, 24, 28_sftp_ablation.py, 32_run_extended_train.py).  The
heavy training of `31_train_with_outputs.py` was originally executed on
the server; the lightweight `23_run_experiments.py` is the local
equivalent that produces the same metrics.

## 4. Datasets and licensing

* The raw ROS bag / CSV recordings are **NOT** redistributed with this
  repository.  Place your own copies under `S:\01_跑车试验数据\...` or
  edit `S_DRIVE` in `05_decode_all.py` to point at your data root.
* The 32 driving segments, 8 scenarios, 65,929 20 Hz control frames and
  the persistent planner-failure regime are documented in the
  accompanying paper (not in this public repo).

## 5. Acknowledgements

* Built on `PyTorch 2.1 + CUDA 12.x` and the `rosbags` Python package.
* Trained on the BISTU shared GPU node (NVIDIA RTX 4090).
* Figures are generated with `matplotlib 3.8` in DejaVu Sans / Times New Roman.

## 6. License

MIT License.