We develop a multi-target multi-branch supernet method, which not only retains the multi-branch structure of HRNet, but also finds the proper location for placing multi-head self-attention module. Our search algorithm is optimized towards multiple objectives (e.g., latency and mIoU) and capable of finding architectures on Pareto frontier with arbitrary number of branches in a single search. We further present a series of HyCTAS that searched for the best hybrid combination of light-weight convolution layers and memory-efficient self-attention layers between branches from different resolutions and fuse to high resolution for both efficiency and effectiveness.

HyCTAS search space

HyCTAS searchable modules

Highlights:

• 1: We design a novel searching framework incorporating with multi-branch space for high resolution representation and genetic-based multi-objective.
• 2: We present a series of HyCTAS that combines a light-weight convolution module to reduce the computation cost while preserving high-resolution information and a memory efficient self-attention module to attend long-range dependencies.
• 3: HyCTAS achieves extremely fast speed, low flops, low parameters and maintains competitive accuracy.

HyCTAS models

• Ubuntu 16.04
• Python 3.7
• CUDA 10.2 (lower versions may work but were not tested)
• NVIDIA GPU (>= 11G graphic memory) + CuDNN v7.3

This repository has been tested on GTX 2080Ti. Configurations (e.g batch size, image patch size) may need to be changed on different platforms.

• Clone this repo:

git clone https://github.com/HyCTAS/HyCTAS.git
cd HyCTAS

• Install dependencies:

bash install.sh

• Download the leftImg8bit_trainvaltest.zip and gtFine_trainvaltest.zip from the Cityscapes.
• Prepare the annotations by using the createTrainIdLabelImgs.py.
• Put the file of image list into where you save the dataset.

• cd HyCTAS/train
• Set the dataset path via ln -s $YOUR_DATA_PATH ../DATASET
• Set the output path via mkdir ../OUTPUT
• Train from scratch

export DETECTRON2_DATASETS="$Your_DATA_PATH"
NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --world_size $NGPUS --seed 12367 --config ../configs/cityscapes/semantic.yaml

We provide training models and logs, which can be downloaded from Google Drive.

cd train

• Download the pretrained weights of the from Google Drive.
• Set config.model_path = $YOUR_MODEL_PATH in semantic.yaml.
• Set config.json_file = $HyCTAS_MODEL in semantic.yaml.
• Start the evaluation process:

CUDA_VISIBLE_DEVICES=0 python test.py

If you find this repository useful, please use the following BibTeX for citation.

@misc{yu2024realtime,
      title={Real-Time Image Segmentation via Hybrid Convolutional-Transformer Architecture Search}, 
      author={Hongyuan Yu and Cheng Wan and Mengchen Liu and Dongdong Chen and Bin Xiao and Xiyang Dai},
      year={2024},
      eprint={2403.10413},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}