README.md

CDP

This repository is for the paper "Bridge the Gap Between Architecture Spaces via A Cross-Domain Predictor".

Preparation

Dependencies

This code is tested on Python 3.6.9, and the dependent packages are listed:

• nasbench (see https://github.com/google-research/nasbench)
• nas_201_api (see https://github.com/D-X-Y/NAS-Bench-201)
• tensorflow (==1.15.0)
• pytorch
• matplotlib
• scipy
• ptflops

Dataset

Two datasets are required:

• NAS-Bench-101
• NAS-Bench-201

Specifically, you can download these two datasets from https://storage.googleapis.com/nasbench/nasbench_only108.tfrecord and https://drive.google.com/open?id=16Y0UwGisiouVRxW-W5hEtbxmcHw_0hF_, separately. And then put these two datasets under the folder path.

The licenses for the datasets can be found in their respective github repositories.

How to use

Search for architectures in DARTS with CDP

python cross_domain_predictor.py

Train the searched architecture on CIFAR-10

python train_cifar10.py --data your_dataset_path

Train the searched architecture on ImageNet

python train_imagenet.py --tmp_data_dir your_dataset_path

Ablation study on dataset tiny DARTS

python read_darts_dataset.py

In addition, you can adjust the parameters following the helps in these files.