GOLD-NAS: Gradual, One-Level, Differentiable

This is an unofficial implementation of GOLD-NAS (https://arxiv.org/abs/2007.03331).

This code is based on the implementation of DARTS.

Requirements

• python 3
• pytorch >= 1.1.0
• torchvision

Results

Please refer to the original paper for complete results.

Usage

Search on CIFAR10

python ./cifar_search/train_search.py \\

Note: in case that you do not have a GPU with 32GB memory, you can reduce the base channel number of search from 36 to 16. We tried and succeeded, but achieved slightly lower accuracy.

Search on ImageNet

We did not implement ImageNet search due to limited computational resource. You can refer to PC-DARTS and embed the ImageNet search code into GOLD-NAS, which is not too difficult.

The evaluation process simply follows that of DARTS.

Here is the evaluation on CIFAR10:

python ./cifar_train/train.py \\