Home

Awesome

BN-NAS: Neural Architecture Search with Batch Normalization

Code for BN-NAS: Neural Architecture Search with Batch Normalization accepted by ICCV2021

<!-- ![introduce image](framework.png) --> <img width="740" height="300" src="framework.png"/>

This project is the re-implementation based on ABS and SPOS.

Requirements

The requirements.txt file lists other Python libraries that this project depends on, and they will be installed using: pip3 install -r requirements.txt

Results

methodArchitectureFLOPsParamsTop-1Supernetea_logretrain model
BNNAS[3, -1, 0, 4, 0, -1, -1, 0, 4, -1, 4, 4, 1, 4, 2, 0, 0, 4, 0, 2, 2]473.5M5.2M75.5supernetea_logretrain model
SPOS[0, -1, 0, 0, 4, -1, -1, 2, 0, -1, 4, 4, 3, 0, 2, 1, 0, 4, 2, 5, 2]468.8M5.8M75.4supernetea_logretrain model

Usage

Step 1: Setup Dataset

Run utils/get_flops_lookup_table.sh to generate flops lookup table which is required in Uniform Sampling.

Step 2: Training supernet

cd BNNAS/supernet
python3 -m torch.distributed.launch --nproc_per_node=8 main.py \
                                    --train_dir YOUR_TRAINDATASET_PATH

Step 3: Search subnets

cd BNNAS/search
cp ../supernet/checkpoint.pth.tar checkpoint.pth.tar
python3 ea.py

Step 3.5 (optional): Show searching result

download the ea_results.pth.tar and put it in BNNAS/search/log

cd BNNAS/search
python3 eval.py

Step 4: Subnet retraining

cd BNNAS/retrain
python3 -m torch.distributed.launch --nproc_per_node=8 train_from_scratch.py \
                            --train_dir $YOUR_TRAINDATASET_PATH --test_dir $YOUR_TESTDATASET_PATH

Usage of SPOS

Step 1: Setup Dataset

Run utils/get_flops_lookup_table.sh to generate flops lookup table which is required in Uniform Sampling.

Step 2: Training supernet

cd SPOS/supernet
python3 -m torch.distributed.launch --nproc_per_node=8 main.py \
                                    --train_dir YOUR_TRAINDATASET_PATH

Step 3: Search subnets

modify the ImageNet Path in SPOS/search/imagenet_dataset.py

cd SPOS/search
cp ../supernet/checkpoint.pth.tar checkpoint.pth.tar
python3 ea.py

Step 3.5 (optional): Show searching result

download the ea_results.pth.tar and put it in SPOS/search/log

cd SPOS/search
python3 eval.py

Step 4: Subnet retraining

cd SPOS/retrain
python3 -m torch.distributed.launch --nproc_per_node=8 train_from_scratch.py \
                            --train_dir $YOUR_TRAINDATASET_PATH --test_dir $YOUR_TESTDATASET_PATH

Thanks

This implementation of BNNAS is based on ABS and SPOS. Please ref to their reposity for more details.

Citation

If you find that this project helps your research, please consider citing our paper:

@inproceedings{chen2021bn,
  title={Bn-nas: Neural architecture search with batch normalization},
  author={Chen, Boyu and Li, Peixia and Li, Baopu and Lin, Chen and Li, Chuming and Sun, Ming and Yan, Junjie and Ouyang, Wanli},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={307--316},
  year={2021}
}