This is the repo of "Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection", which is accepted by AAAI 2024. We would publish the pre-print version (containing the supplementary, intending for a better understanding) and the source code.
We built this system on other excellent projects:
- mmdetection: https://github.com/open-mmlab/mmdetection
- condlane: https://github.com/aliyun/conditional-lane-detection
- ganet: https://github.com/Wolfwjs/GANet
- pytorch-auto-drive: https://github.com/voldemortX/pytorch-auto-drive
- qdrop: https://github.com/wimh966/QDrop
- install mmdet
cd mmdet ; python setup.py develop
- prepare the TuSimple and CurveLane datasets according to PAD. It is worthy note that the BezierLaneNet needs extra data-preprocessing as it adopted a curve-based modeling of lanes. The model checkpoints could be find at PAD, CondlaneNet, GANet.
Also remember to compile the unified evaluate tool on evaluate/culane/ by
make
- run
we provide scripts to run different quantization settings in the dir scripts and the network and quantization configs are located at dir config.
To run PTQ on PAD[https://github.com/voldemortX/pytorch-auto-drive] models:
python tools/pad/main_lanedet_quant.py \
--config $network_cfg \
--quant_config $quant_cfg \
--test \
--output_prefix $output_dir \
2>&1 | tee -a $log_file_name
To run PTQ on CondlaneNet and GANet:
python tools/mmdet/culane_ganet_quant.py \
$network_cfg \
$quant_cfg \
$ckpt_pth_file \
--result_dst $output_dir \
--launcher none 2>&1 | tee -a $log_file_name
NOTE: every pair of (dataset, network) gets a single scripts in the tools/.
An arxiv version is on going...