This is an official implementation for NEAS presented in CVPR 2021.

To set up the enviroment you can easily run the following command:

git clone https://github.com/researchmm/NEAS.git
cd NEAS
conda create -n NEAS python=3.6
conda activate NEAS
sh ./install.sh
# (required) install apex to accelerate the training, a little bit faster than pytorch DistributedDataParallel
cd lib
git clone https://github.com/NVIDIA/apex.git
python ./apex/setup.py install --cpp_ext --cuda_ext

You need to first download the ImageNet-2012 to the folder ./data/imagenet and move the validation set to the subfolder ./data/imagenet/val. To move the validation set, you cloud use the following script: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh

The directory structure is the standard layout as following.

/path/to/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class/2
      img4.jpeg

For evaluation, we provide the checkpoints of our models in Google Drive.

After downloading the models, you can do the evaluation following the description in Quick Start - Test).

Model download links:

We provide test code of NEAS as follows.

To test our trained models, you need to put the downloaded model in PATH_TO_CKP (the default path is ./CKP in root directory.). After that you need to specify the model path in the corresponding config file by changing the intitial-checkpoint argument in ./configs/subnets/[SELECTED_MODEL_SIZE].yaml.

Then, you could use the following command to test the model.

sh ./tools/distributed_test.sh ./configs/subnets/[SELECTED_MODEL_SIZE].yaml

The test result will be saved in ./experiments. You can also add [--output OUTPUT_PATH] in ./tools/distribution_test.sh to specify a path for it as well.

• Test code
• Retrain code
• Search code

@article{NEAS,
  title={One-Shot Neural Ensemble Architecture Search by Diversity-Guided Search Space Shrinking},
  author={Chen, Minghao and Peng, Houwen and Fu, Jianlong and Ling, Haibin},
  journal={arXiv preprint arXiv:2104.00597},
  year={2021}
}