/DARTS_evaluation

Train from scratch for architectures from DARTS space

Primary LanguagePythonMIT LicenseMIT

Evaluation Tool for architectures from DARTS search space

This a slimmed repository for evaluating architectures searched from DARTS search space (originally from NASNET search space) on CIFAR-10/100 and ImageNet, i.e., training an architecutre from scratch.

This repository is also used to evaluate the searched architectures by TENAS and TEGNAS. Genotypes of our searched architectures are listed in genotypes.py

Usage:

Please put your searched genotype inside genotypes.py.

CIFAR-10

# CIFAR-10 training
CUDA_VISIBLE_DEVICES=0 python train.py --cutout --auxiliary --data /raid/ --arch TENAS_cifar10 --batch_size 96 --epoch 800

ImageNet

The settings batch_size = 768 and learning_rate = 0.5 by default are designed for 8-gpu training.

# ImageNet distributed training
python train_imagenet_dist.py --batch_size 768 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --data /raid/imagenet --arch TENAS_imagenet\
--save_root ./