SinglePathOneShotNAS-Gluon

Gloun Version of SinglePathOneShotNAS

Official Pytorch Implementation

Single Path One-Shot Neural Architecture Search with Uniform Sampling

Prepare

Packages

pip install mxnet-cu101mkl(depend on you CUDA version)

pip install gluoncv

pip install mxboard

pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/cuda/10.0 nvidia-dali

pip install horovod

pip install mxop(Optional)

Dataset

cd $ROOT && mkdir data

cd data && ln -s path/to/imagenet_image_dir/ ./ or cd data && ln -s path/to/imagenet_rec_dir/ ./

SuperNet

Train supernet with:sh train_supernet.sh

Remark: Make sure your dataset path is correct before training.

Search

Search subnet with flops/params limits:sh search.sh

Remark: Change resume-params in search.py or search.sh, also you should make sure your dataset path is correct before searching, change flops/params limit if you like.

Evaluate

Retrain the best subnet:sh eval.sh

Remark: Make sure your dataset path is correct before retraining.

Reference

https://github.com/megvii-model/SinglePathOneShot

https://gluon-cv.mxnet.io/model_zoo/classification.html

https://github.com/CanyonWind/Single-Path-One-Shot-NAS-MXNet

https://github.com/hey-yahei/OpSummary.MXNet