This repository is an official code implementation of FGP, a flow-based generative pre-training method for neural architecture encoding.

• Title: Learning to Flow from Generative Pretext Tasks for Neural Architecture Encoding.
• Authors: Sunwoo Kim, Hyunjin Hwang, Kijung Shin
• Venue: NeurIPS 2025

We use three datasets: NAS-Bench-101 [1], NAS-Bench-201 [2], and NAS-Bench-301 [3].

In the following link, we provide our pre-processed version of these datasets. https://drive.google.com/drive/folders/163H_FbKeeng0rWVtKHiYiEY05qdK6GQX?usp=sharing

1. Please make the folder named ./dataset in the same hierarchy with main.py
2. Then, please put the datasets in ./dataset folder.

Refer to the following notebook:

obtaining_flow_surrogate.ipynb

One can reproduce the performance prediction results by running the following code:

python3 main.py -dname nb101 -device cuda:0 -train_ratio 0.01 -proj_layer 2 -proj_dim1 128 -proj_dim2 64 -ssl_lr 0.001 -wdecay 1e-6 -lamda1 0.5 -lamda2 0.5 -enc gatedgcn

Best hyperparameter configurations are in fgp_hyperparameters.pickle.

Each argument corresponds to each hyperparameter, which is as follows:

• -dname is a str that indicates the data name, which should be one of {nb101, nb201, nb301}.
• -device is a str that indicates the data type that indicates GPU device number, such as cuda:0.
• train_ratio is a float that indicates the ratio of the training set to be used for model fine-tuning, which should be within $(0,1]$, such as 0.01.
• proj_layer is a int that indicates the number of layers of the projection head, such as 2.
• proj_dim1 is a int that indicates the hidden dimension of the surrogate projection head, such as 128.
• proj_dim2 is a int that indicates the hidden dimension of the zero-cost proxy projection head, such as 64.
• ssl_lr is a float that indicates the pre-training learning rate, such as 1e-3.
• wdecay is a float that indicates the pre-training weight decay, such as 1e-6.
• lamda1 is a float that indicates the loss coefficient of the surrogate reconstruction loss, such as 0.5.
• lamda2 is a float that indicates the loss coefficient of the zero-cost proxy prediction loss, such as 0.5.
• enc is a str that indicates the neural architecture encoder type, which should be one of {gatedgcn, gin, flowerformer}.

• Supervised training code is from (Hwang et al., FlowerFormer: Empowering Neural Architecture Encoding using a Flow-aware Graph Transformer, In CVPR 2024).
• [1] Ying et al., NAS-Bench-101: Towards Reproducible Neural Architecture Search, In ICML 2019.
• [2] Dong et al., NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search, In ICLR 2020.
• [3] Zela et al., Surrogate NAS Benchmarks: Going Beyond the Limited Search Spaces of Tabular NAS Benchmarks, In ICLR 2022.