Task: In-Cabin User Command Understanding (UCU), workshop in WACV2024
Here is our solution code. Please check the report for more detail.
Here we introduced how to setup and the way to downloaded their model:
-
Send request to their form Apply here in Meta official page and you will receive an email with some details.
-
Dependencies:
sudo apt install wget ucommon-utils
-
run
./scripts/llama/download.sh
to download Llama models. -
mamba create --name llc python=3.8 && mamba activate llc && pip install -r requirements.txt
-
Run the example:
torchrun --nproc_per_node 1 scripts/main_codellama.py
copy your OPENAI_API_KEY and save it in .env
.
OPENAI_API_KEY='xxxx'
install below
pip install openai
pip install -U python-dotenv
Run the example:
python scripts/main_gpt.py --provide_few_shots True --step_by_step True
You will have a result .json
file finally. Then run the eval.py
For each task-level accuracy:
python scripts/eval.py -g assets/ucu.csv -e assets/result/test.json
Here is the demo output:
Evaluating assets/result/gpt-4_best.json ...
| Task | Accuracy |
|---------------------+------------|
| Perception | 0.931756 |
| In-cabin monitoring | 0.748863 |
| Localization | 0.915378 |
| Vehicle control | 0.88626 |
| Entertainment | 0.944495 |
| Personal data | 0.859873 |
| Network access | 0.919927 |
| Traffic laws | 0.915378 |
| Overall | 0.890241 |
Here is official evaluate with -o
:
python3 scripts/llvm_ad/official_eval.py -o -g assets/ucu.csv -e assets/result/gpt4_best.csv
We attach the raw .json
output files (with explainations and output), and its corresponding .cvs
files (binary output only) under assets/result
folder.
Here is demo output:
Since the input file is .json, we save the prediction to .csv file at assets/result/gpt-4_best.csv
Evaluating assets/result/gpt-4_best.json ...
Following is the evaluation result in official way:
Command-level acc: 0.38034576888080074
Question-level acc: 0.8902411282984531
This work was funded by Vinnova, Sweden (research grant). The computations were enabled by the supercomputing resource Berzelius provided by National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg foundation, Sweden.
This implementation is based on codes from several repositories. Thanks for these authors who kindly open-sourcing their work to the community. Please see our paper reference part to get more information.
@inproceedings{yang2024human,
title={Human-centric autonomous systems with llms for user command reasoning},
author={Yang, Yi and Zhang, Qingwen and Li, Ci and Marta, Daniel Sim{\~o}es and Batool, Nazre and Folkesson, John},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={988--994},
year={2024}
}