Neural Architecture Transfer

Code accompanying the paper.

Neural Architecture Transfer

Zhichao Lu, Gautam Sreekumar, Erik Goodman, Wolfgang Banzhaf, Kalyanmoy Deb, and Vishnu Boddeti

arXiv:2005.05859

overview

Requirements

Python >= 3.7.x, PyTorch >= 1.4.0, timm >= 0.1.18

ImageNet Classification

imagenet

python evaluator.py --data /path/to/dataset --model subnets/imagenet/NAT-{M1,M2,M3,M4}/net.config

Architecture Transfer

transfer

python evaluator.py \
  --data /path/to/dataset \
  --dataset {aircraft,cars,cifar10,cifar100,cinic10,dtd,flowers102,food101,pets,stl10} \
  --model subnets/{dataset}/net-img@{xxx}-flops@{xxx}-top1@{xx.x}/net.config

Scalability to Objectives

Tri-Objective

tri-obj

12-Objective

12-obj

Architecture Visualization

visualization

Acknowledgement

Codes are modified from OnceForAll and pytorch-image-models