Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets

MetaD2A

This is the official PyTorch implementation for the paper Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets (ICLR 2021) : https://openreview.net/forum?id=rkQuFUmUOg3.

Abstract

Despite the success of recent Neural Architecture Search (NAS) methods on various tasks which have shown to output networks that largely outperform human-designed networks, conventional NAS methods have mostly tackled the optimization of searching for the network architecture for a single task (dataset), which does not generalize well across multiple tasks (datasets). Moreover, since such task-specific methods search for a neural architecture from scratch for every given task, they incur a large computational cost, which is problematic when the time and monetary budget are limited. In this paper, we propose an efficient NAS framework that is trained once on a database consisting of datasets and pretrained networks and can rapidly search a neural architecture for a novel dataset. The proposed MetaD2A (Meta Dataset-to-Architecture) model can stochastically generate graphs (architectures) from a given set (dataset) via a cross-modal latent space learned with amortized meta-learning. Moreover, we also propose a meta-performance predictor to estimate and select the best architecture without direct training on target datasets. The experimental results demonstrate that our model meta-learned on subsets of ImageNet-1K and architectures from NAS-Bench 201 search space successfully generalizes to multiple benchmark datasets including CIFAR-10 and CIFAR-100, with an average search time of 33 GPU seconds. Even under a large search space, MetaD2A is 5.5K times faster than NSGANetV2, a transferable NAS method, with comparable performance. We believe that the MetaD2A proposes a new research direction for rapid NAS as well as ways to utilize the knowledge from rich databases of datasets and architectures accumulated over the past years.

Framework of MetaD2A Model

Prerequisites

• Python 3.6 (Anaconda)
• PyTorch 1.6.0
• CUDA 10.2
• python-igraph==0.8.2
• tqdm==4.50.2
• torchvision==0.7.0
• python-igraph==0.8.2
• nas-bench-201==1.3
• scipy==1.5.2

If you are not familiar with preparing conda environment, please follow the below instructions

$ conda create --name metad2a python=3.6
$ conda activate metad2a
$ conda install pytorch==1.6.0 torchvision cudatoolkit=10.2 -c pytorch
$ pip install nas-bench-201
$ conda install -c conda-forge tqdm
$ conda install -c conda-forge python-igraph
$ pip install scipy

And for data preprocessing,

$ pip install requests

Hardware Spec used for experiments of the paper

• GPU: A single Nvidia GeForce RTX 2080Ti
• CPU: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz

NAS-Bench-201

Go to the folder for NAS-Bench-201 experiments (i.e. MetaD2A_nas_bench_201)

$ cd MetaD2A_nas_bench_201

Data Preparation

To download preprocessed data files, run get_files/get_preprocessed_data.py:

$ python get_files/get_preprocessed_data.py

It will take some time to download and preprocess each dataset.

To download MNIST, Pets and Aircraft Datasets, run get_files/get_{DATASET}.py

$ python get_files/get_mnist.py
$ python get_files/get_aircraft.py
$ python get_files/get_pets.py

Other datasets such as Cifar10, Cifar100, SVHN will be automatically downloaded when you load dataloader by torchvision.

If you want to use your own dataset, please first make your own preprocessed data,
by modifying process_dataset.py .

$ process_dataset.py

MetaD2A Evaluation (Meta-Test)

You can download trained checkpoint files for generator and predictor

$ python get_files/get_checkpoint.py
$ python get_files/get_predictor_checkpoint.py

1. Evaluation on Cifar10 and Cifar100

By set --data-name as the name of dataset (i.e. cifar10, cifar100),
you can evaluate the specific dataset only

# Meta-testing for generator 
$ python main.py --gpu 0 --model generator --hs 56 --nz 56 --test --load-epoch 400 --num-gen-arch 500 --data-name {DATASET_NAME}

After neural architecture generation is completed,
meta-performance predictor selects high-performing architectures among the candidates

# Meta-testing for predictor
$ python main.py --gpu 0 --model predictor --hs 512 --nz 56 --test --num-gen-arch 500 --data-name {DATASET_NAME}

2. Evaluation on Other Datasets

By set --data-name as the name of dataset (i.e. mnist, svhn, aircraft, pets),
you can evaluate the specific dataset only

# Meta-testing for generator
$ python main.py --gpu 0 --model generator --hs 56 --nz 56 --test --load-epoch 400 --num-gen-arch 50 --data-name {DATASET_NAME}

After neural architecture generation is completed,
meta-performance predictor selects high-performing architectures among the candidates

# Meta-testing for predictor
$ python main.py --gpu 0 --model predictor --hs 512 --nz 56 --test --num-gen-arch 50 --data-name {DATASET_NAME}

Meta-Training MetaD2A Model

You can train the generator and predictor as follows

# Meta-training for generator
$ python main.py --gpu 0 --model generator --hs 56 --nz 56 
                 
# Meta-training for predictor
$ python main.py --gpu 0 --model predictor --hs 512 --nz 56 

Results

The results of training architectures which are searched by meta-trained MetaD2A model for each dataset

Accuracy

Search Time (GPU Sec)

MobileNetV3 Search Space

Go to the folder for MobileNetV3 Search Space experiments (i.e. MetaD2A_mobilenetV3)

$ cd MetaD2A_mobilenetV3

And follow README.md written for experiments of MobileNetV3 Search Space

GitHub

https://github.com/HayeonLee/MetaD2A