EfficientNet-eLite: Extremely Lightweight and Efficient CNN Models for Edge Devices by Network Candidate Search

Pytorch implementation for paper EfficientNet-eLite

We purpose a family of Extremely lightweight CNN models by EfficientNet to be effectively scaled down. The following image illustrates the performance of parameter usage and the top-1 accuracy on ImageNet dataset. More details can be found at paper EfficientNet-eLite

Environment

Anaconda 4.7.12

• Python 3.6.9

• pytorch 1.2.0, torchvision 0.4.0, cuda10

• git+https://github.com/wbaek/theconf@de32022f8c0651a043dc812d17194cdfd62066e8

• git+https://github.com/ildoonet/pytorch-gradual-warmup-lr.git@08f7d5e

• git+https://github.com/ildoonet/pystopwatch2.git

• git+https://github.com/hyperopt/hyperopt.git

• pretrainedmodels

• tqdm

• tensorboardx

• sklearn

• ray

• matplotlib

• psutil

• requests

Start training a model

Example for EfficientNet-eLite 9

$ bash run.sh

source ~/miniconda3/etc/profile.d/conda.sh 

conda activate 'name of your conda environment'

python3 FastAutoAugment/train.py -c confs/efficientnet_l9.yaml --aug fa_reduced_imagenet  --dataroot 'Path of ImageNet on server'

Note :

  1. Specify -c confs/efficientnet_l8.yaml -c confs/efficientnet_l7.yaml ...... to begin the other training of EfficientNet-eLite family

  2. --aug fa_reduced_imagenet to select the data augmentation policy (implementation is from Fast Autoaugmentation)

  3. --dataroot should be configured as the path root of ImageNet dataset with the subfolder consisting 'train' and 'val'.

  In the inside folder of train and val, each subfolder has the name of the label and the organization is the same for using ImageFolder from torchvision