Encodings for Prediction-based Neural Architecture Search

Authors: Yash Akhauri, Mohamed S Abdelfattah

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our analysis draws from experiments conducted on over 1.5 million neural network architectures on NAS spaces such as NASBench-101 (NB101), NB201, NB301, Network Design Spaces (NDS), and Trans NASBench101. Building on our study, we present our predictor FLAN: Flow Attention for NAS.
Researcher Affiliation Academia Yash Akhauri 1 Mohamed S. Abdelfattah 1 1Cornell University, New York, USA. Correspondence to: Yash Akhauri <ya255@cornell.edu>.
Pseudocode No The paper describes its methods and uses equations, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Our implementation and encodings for all neural networks are open-sourced at https://github.com/abdelfattahlab/flan nas.
Open Datasets Yes NASBench-101(Ying et al., 2019) and NASBench-201(Dong & Yang, 2020) are search spaces based on cells, comprising 423,624 and 15,625 architectures respectively. NASBench-101 undergoes training on CIFAR-10, whereas NASBench-201 is trained on CIFAR-10, CIFAR100, and Image Net16-120. NASBench-301(Zela et al., 2020) serves as a surrogate NAS benchmark, containing a total of 1018 architectures. Trans NAS-Bench-101(Duan et al., 2021) stands as a NAS benchmark that includes a micro (cell-based) search space with 4096 architectures and a macro search space embracing 3256 architectures.
Dataset Splits No While the paper specifies training and testing sample counts and refers to 'validation accuracy' from prior work's calculations (Figure 15), it does not explicitly provide details or percentages for a validation dataset split used in its own experimental setup.
Hardware Specification No The paper mentions running experiments on a 'single GPU' and a 'consumer GPU' but does not provide specific model numbers (e.g., NVIDIA A100, RTX 3090), processor types, or memory details for the hardware used.
Software Dependencies No Table 12 lists hyperparameters and training settings, such as Learning Rate and Batch Size. However, it does not specify software dependencies like programming language versions (e.g., Python 3.8) or library versions (e.g., PyTorch 1.9).
Experiment Setup Yes Table 12: Hyperparameters used in main table experiments. Learning Rate 0.001 Weight Decay 0.00001 Number of Epochs 150 Batch Size 8