MOTE-NAS: Multi-Objective Training-based Estimate for Efficient Neural Architecture Search

This project provides a few-cost estimate for Neural Architecture Search.

1. This code covers the core function of the proposed MOTE.
2. File "example.py" and "gen_mote.py" is the main program, which records the code for MOTE generation.
3. Files "meta_models.py", "reduced_data.py" respectively correspond to the core methods mentioned in the paper - Reduced Architecture(RA) and Reduced Data(RD).

The paper has been accepted by NeurIPS 2024.

Paper Link: https://reurl.cc/KdMalq

Requirements

• Python 3
• TensorFlow 2
• OpenCV
• Numpy

How to Get MOTE for an Architecture(cell_pth) ?

from nas_prcss import CellPth2Cell
from gen_mote import GetProxyC100TrainData, CellPth2MOTE
train_x,train_y=GetProxyC100TrainData(labels_len=10)
CellPth2MOTE(train_x,train_y,cell_pth,cell_type="nas201",proxy_labels_len=10)
cell_dict=CellPth2Cell(cell_pth)
print(cell_dict["mote"])

How to Get Correlation of MOTE Ranking and Actual Ranking ?

from nas_prcss import SamplingCellPths,FilteringByDirtyBit
from eval_cell import CellPths2Psp
cells_dir="data/nasbench201_img16-10"
gt_key="test_accuracy_200"
cell_pths=SamplingCellPths(cells_dir,shuffle=True)
psp=CellPths2Psp(cell_pths,gt_key,"mote")
print(psp)

Fully Dataset

The NASBench-101 data set is too large, here only the NASBench-201, we will provide NB101 ASAP.

Our Data Format

coming soon.

The Comparison of Search Efficiency on NASBench-201

For detailed report, please refer to our paper.

TODOs

• Provide NB101.
• Provide more complete code.