RBFleX-NAS : Training-Free Neural Architecture Search Using Radial Basis Function Kernel and Hyperparameter Detection

Training-Free Neural Architecture Search (NAS) Using Radial Basis Function (RBF) Kernel

Demo

🟨 Download the paper

RBFleX-NAS: Training-Free Neural Architecture Search Using Radial Basis Function Kernel and Hyperparameter Detection Tomomasa Yamasaki; Zhehui Wang; Tao Luo; Niangjun Chen; Bo Wang IEEE Transactions on Neural Networks and Learning Systems (TNNLS)

🟨 What is RBFleX-NAS

Neural Architecture Search (NAS) is an automated technique to design optimal neural network architectures for a specific workload. Conventionally, evaluating candidate networks in NAS involves extensive training, which requires significant time and computational resources. To address this, training-free NAS has been proposed to expedite network evaluation with minimal search time. However, state-of-the-art training-free NAS algorithms struggle to precisely distinguish well-performing networks from poorly-performing networks, resulting in inaccurate performance predictions and consequently sub-optimal top-1 network accuracy. Moreover, they are less effective in activation function exploration. To tackle the challenges, this paper proposes RBFleX-NAS, a novel training-free NAS framework that accounts for both activation outputs and input features of the last layer with a Radial Basis Function (RBF) kernel. We also present a detection algorithm to identify optimal hyperparameters using the obtained activation outputs and input feature maps. We verify the efficacy of RBFleX-NAS over a variety of NAS benchmarks. RBFleX-NAS significantly outperforms state-of-the-art training-free NAS methods in terms of top-1 accuracy, achieving this with short search time in NAS-Bench-201 and NAS-Bench-SSS. In addition, it demonstrates higher Kendall correlation compared to layer-based training-free NAS algorithms. Furthermore, we propose NAFBee, a new activation design space that extends the activation type to encompass various commonly used functions. In this extended design space, RBFleX-NAS demonstrates its superiority by accurately identifying the best-performing network during activation function search, providing a significant advantage over other NAS algorithms.

🟨 Environmental Requirements

For RBFleX-NAS

• Python version 3
• numpy
• torchvision
• torch
• scipy
• nats_bench

🟨 Download NAS Benchmarks

Our program works for NAS-Bench-201 (NATS-Bench-TSS), NATS-Bench-SSS, Network Design Space, and TransNAS-Bench-101. If you want to apply our NAS algorithm, edit our program to meet other NAS benchmarks.

NATS-Bench

To download the latest version, please visit the GitHub page of NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size. You can download the benchmark tar file of NATS-tss-v10-3ffb9-simple.tar and NATS-sss-v10-50262-simple.tar from Google drive. After downloading these benchmark tar files, please store "NATS-tss-v10-3ffb9-simple.tar" on ./designspace/NAS-Bench-201/ and "NATS-sss-v10-50262-simple.tar" on ./designspace/NATS-Bench-SSS/, respectively. And then, you may uncompress them by running tar xvf NATS-tss-v10-3ffb9-simple.tar and tar xvf NATS-sss-v1_0-50262-simple.tar.

Network Design Space (NDS)

To download NDS benchmark json files, please visit the GitHub page of On Network Design Spaces for Visual Recognition. You can download the benchmark json files from this link. After downloading these json files, please store all of json files on ./designsoace/NDS/

TransNAS-Bench-101

TransNAS-Bench-101 is put on this GitHub. You may not download TransNAS-Bench-101. To know the detail of TransNAS-Bench-101 benchmark, please visit the GitHub page of TransNAS-Bench-101: Improving Transferability and Generalizability of Cross-Task Neural Architecture Search. TransNAS-Bench-101 is stored on ./designsoace/TransMacro/ for Macro-level design space and ./designsoace/Trans_Micro/ for cell-level design space.

NAFBee: Neural Network Activation Function Benchmark

This is a benchmark for networks with a variety of activation functions based on VGG-19 and BERT. NAFBee provides network information and accuracy. Users can obtain the accuracy without training. To download NAFBee, please visit the GitHub Page of NAFBee. NAFBee is stored on ./designsoace/NAFBee/

🟨 Download image dataset

Our program works with three major datasets: CIFAR-10, CIFAR-100, ImageNet, Taskonomy, and SST-2.

CIFAR-10 for NATS-Bench, NDS, and NAFBee(VGG-19)

This dataset is automatically downloaded on ./dataset/cifar10 by PyTorch after running our program. If you already have CIFAR-10 dataset in your environment, you may set your dataset path on root on each main program as follows.

CIFAR-100 for NATS-Bench

This dataset is automatically downloaded on ./dataset/cifar100 by PyTorch after running our program. If you already have CIFAR-100 dataset in your environment, you may set your dataset path on root on each main program as follows.

ImageNet or ImageNet16-120 for NATS-Bench and NDS

You may download a training set of ImageNet from here. Before downloading, you should create an account on the website. After downloading it, please store the dataset in ./dataset/ImageNet. If you use ImageNet16-120 for NATS-Bench, please store the dataset in ./dataset/ImageNet16. NATS-Bench GitHub explains the ImageNet16-120 dataset.

Taskonomy for TransNAS-Bench-101

You may download Taskonomy RGB images as the dataset from website, and then put them in ./dataset/TASKONOMY/rgb.

SST-2 for NAFBee(BERT)

You may download SST-2 dataset in ./dataset/SST2.

🟨 How to Run

1. Download NAS Benchmarks

put them on ./designspace/NAS-Bench-201/, ./designspace/NATS-Bench-SSS/, ./designsoace/NDS/, ./designsoace/TransMacro/, ./designsoace/TransMicro/, or ./designsoace/NAFBee/.

2. Set parameters

For NAS-Bench-201 or NATS-Bench-SSS, please set batch size for RBFleX-NAS, dataset, number of trials, and a number of networks selected randomly from benchmark

# ==============================================
GLOBAL VARIABLE: Batch size for RBFleX-NAS
==============================================
batchsizeNE = 3
==============================================
GLOBAL VARIABLE: Experiment for RBFleX-NAS
- cifar10
- cifar100
- ImageNet16-120
==============================================
dataset = 'cifar10'
==============================================
GLOBAL VARIABLE: Experiment for RBFleX-NAS
maxtrials: a number of trials
Num_Networks: a number of networks selected randomly from benchmark
==============================================
maxtrials = 10
Num_Networks = 1000

For TransNAS-Bench-101 (Micro and Macro), Please set DESIGN SPACE CONFIGURATION and batch size for RBFleX-NAS

# ============================================== 
GLOBAL VARIABLE: DESIGN SPACE CONFIGURATION
 
- configmacstransmicrosegmentsemantic
 
- configmacstransmicroautoencoder
 
- configmacstransmicroclassobject
 
- configmacstransmicroclassscene
 
- configmacstransmicrojigsaw
 
- configmacstransmicronormal
 
- configmacstransmicroroomlayout
 
==============================================
 CONFIGPATH = "configmacs_transmicrosegmentsemantic"

==============================================
GLOBAL VARIABLE: Batch size for RBFleX-NAS
==============================================
batchsizeNE = 3

For NAFBee(BERT), Please set batch size for RBFleX-NAS

# ============================================== 
GLOBAL VARIABLE: Batch size for RBFleX-NAS
 
==============================================
 batchsizeNE = 3

3. Run an experiment to reproduce a result

The following programs use the fixed hyperparameters for RBF kernels. (2.68690173088039e-12 for Activation, 1.02460061284506e-11 for final input)

For NAS-Bench-201 or NATS-Bench-SSS,

python RBFleX_NAS-Bench-201.py
or
python RBFleX_NATS-Bench-SSS.py

For TransNAS-Bench-101 (Micro and Macro),

python RBFleXTransNASMacro.py

or
python RBFleXTransNASMicro.py

This program runs with the hyperparameter detection algorithm (HDA) proposed in the RBFleX-NAS paper.

For NAFBee(BERT),

python RBFleXNAFBeeBERT.py

🟨 Hyperparameter Detection Algorithm (HDA)

RBFleX-NAS uses RBF kernels to evaluate the similarities of activation outputs and final inputs. However, an RBF kernel is sensitive to the hyperparameter GAMMA. Therefore, RBFleX-NAS detects the optimal hyperparameters based on measured features. You may run the hyperparameter detection algorithm (HDA) like

python HDA.py

Note: HDA.py is an example for NATS-Bench-SSS. If you want to leverage HDA to other design space, please set these configurations. NGAMMA is a number of networks to detect hyperparameter for an RBF kernel. searchspace is a design space for ./designspace/NAS-Bench-201/, ./designspace/NATS-Bench-SSS/, ./designsoace/NDS/, ./designsoace/TransMacro/, ./designsoace/Trans_Micro/, or ./designsoace/NAFBee/. Please refer these GitHub page to understand how to define search space.

# ============================================== 
GLOBAL VARIABLE: 
 
Batch size for RBFleX-NAS
 
N_GAMMA: Number of networks to detect hyperparameter for RBF kernel
 
==============================================
 batchsizeNE = 3 N_GAMMA = 10

==============================================
GLOBAL VARIABLE: Experiment for RBFleX-NAS
- cifar10
- cifar100
- ImageNet16-120
==============================================
dataset = 'cifar10'
==============================================
GLOBAL VARIABLE: create a searchspace
This example is NATS-Bench-SSS
==============================================
benchmarkroot = "./designspace/NATS-Bench-SSS/NATS-sss-v10-50262-simple"
NAS Benchmark
print('Loading...NAT Bench '+"sss")
searchspace = create(benchmarkroot, "sss", fastmode=True, verbose=False)
==============================================
GLOBAL VARIABLE: create a dataloader
This example is cifar-10
==============================================
img_root = "./dataset"
print('==> Preparing data..')
transform_train = transforms.Compose([
    transforms.RandomCrop(32, padding=4),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
imgset = torchvision.datasets.CIFAR10(
        root=imgroot+'/cifar10', train=True, download=True, transform=transformtrain)
img_loader = torch.utils.data.DataLoader(
        imgset, batchsize=batchsizeNE, shuffle=True, numworkers=1, pin_memory=True)

🟨 Citing RBFleX-NAS

If you use RBFleX, please cite the following paper:

@ARTICLE{10959729,
  author={Yamasaki, Tomomasa and Wang, Zhehui and Luo, Tao and Chen, Niangjun and Wang, Bo},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={RBFleX-NAS: Training-Free Neural Architecture Search Using Radial Basis Function Kernel and Hyperparameter Detection}, 
  year={2025},
  volume={36},
  number={6},
  pages={10057-10071},
  keywords={Accuracy;Training;Vectors;Kernel;Feature extraction;Correlation;Computer architecture;Benchmark testing;Predictive models;Neural architecture search;Activation;neural architecture search (NAS);radial basis function (RBF);training-free},
  doi={10.1109/TNNLS.2025.3552693}}

🟨 Licence

MIT Licence

Contact