Homogeneous Architecture Augmentation for Neural Predictor
This repository is for the paper "Homogeneous Architecture Augmentation for Neural Predictor" which is accepted by ICCV 2021.
Introduction
The codes have been tested on Python 3.6.
Dependent packages:
- 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)
- scikit-learn
- matplotlib
- scipy
The pkl folder saves the fixed training set and fixed test set.
If you would like to check other training data from NAS-Bench-101, please download the NAS-Bech-101 subset of the dataset with only models trained at 108 epochs: https://storage.googleapis.com/nasbench/nasbench_only108.tfrecord. (More details are in https://github.com/google-research/nasbench, and you may be required to install additional dependencies like TensorFlow.) Then put the file nasbench_only108.tfrecord under the path folder. Finally, carefully delete the corresponding files in pkl folder.
How to use
Demo0.py is for Table 1 and Figure 5. If you want to see NPNAS + HA, run the neuralpredictor.pytorch/train.py (--arch_aug is an argument).
GAon201.py is for Table 3.
Demo1.py is for Table 4.
GAon101/random_forest_Surrogate.py is the Demo2, which is for Table 2. And this must need to download the file nasbench_only108.tfrecord. You can find the results in GAon101/pops_log.
Demo3.py is for Table 5.
Demo5.py is for Figure 7. If you want to test randomly, please delete the pkl/num_creations.pkl.
You can run these scripts to get the results reported in paper. You can change parameter settings following the annotations between codes.