AutoLC

[ICPR 2022] AutoLC: Search Lightweight and Top-Performing Architecture for Remote Sensing Image Land-Cover Classification. Including implementations of AutoDeepLab, Deeplab, Deeplabv3, Deeplabv3+, U-Net, FCN, PSPNet on LoveDA, Cityscapes and Floodnet data sets.

File description

mypath.py: path to each dataset (including loveda, cityscape, floodnet)
autodeeplab.py: architecture Level search space of AutoDeepLab
cell_level_search.py: cell-level search space of AutoDeepLab
genotypes.py: module candidates of cell-level search space
modeling, operations.py: all basic models and operations included in our paper
decode.py: decode searched architecture
decoding_formulas.py: concrete implement for decoding architecture
retrain_model: the model we proposed with lightweight encoder and decoder
search.py: neural architecture search
retrain.py: retrain the network after decoding
test.py: get predictions on test set
stats: get parameters and computational consumption of different models

Training Procedure

All together there are 3 stages:

Architecture Search - Here you will train one large relaxed architecture that is meant to represent many discreet smaller architectures woven together. See search.py.
Decode - Once you've finished the architecture search, load your large relaxed architecture and decode it to find your optimal architecture. See decode.py: decode searched architecture.
Re-train - Once you have a decoded and poses a final description of your optimal model, use it to build and train your new optimal model. See retrain.py.

Hardware Requirement

For architecture search with config in config_utils/search_args.py, you need at least an 12G GPU.

Reproduce

Search the lightweight encoder

bash scripts/search_loveda.sh

Decode to get the encoder

bash scripts/decode.sh

Retrain the decoded architecture with lightweight decoder

bash scripts/retrain.sh

Get prediction on the test set

bash scripts/test.sh

Requirements

Pytorch
Python 3
tensorboardX
torchvision
tqdm
numpy
pandas
apex

Citation

If you use AutoLC in your research, please cite our ICPR2022 paper.

@inproceedings{DBLP:conf/icpr/ZhengWMZ22,
  author    = {Chenyu Zheng and
               Junjue Wang and
               Ailong Ma and
               Yanfei Zhong},
  title     = {AutoLC: Search Lightweight and Top-Performing Architecture for Remote
               Sensing Image Land-Cover Classification},
  booktitle = {26th International Conference on Pattern Recognition, {ICPR} 2022,
               Montreal, QC, Canada, August 21-25, 2022},
  pages     = {324--330},
  publisher = {{IEEE}},
  year      = {2022},
  url       = {https://doi.org/10.1109/ICPR56361.2022.9956372},
  doi       = {10.1109/ICPR56361.2022.9956372},
  timestamp = {Thu, 01 Dec 2022 15:50:19 +0100},
  biburl    = {https://dblp.org/rec/conf/icpr/ZhengWMZ22.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

References

[1] : Thanks for NoamRosenbergs autodeeplab model implemention