Accelerating Neural Architecture Search via Proxy Data

Authors: Byunggook Na, Jisoo Mok, Hyeokjun Choe, Sungroh Yoon

IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To empirically demonstrate the effectiveness, we conduct thorough experiments across diverse datasets, search spaces, and NAS algorithms.
Researcher Affiliation Academia 1 Department of Electrical and Computer Engineering, Seoul National University, Seoul, South Korea 2 AIIS, ASRI, INMC, and Interdisciplinary Program in AI, Seoul National University, Seoul, South Korea
Pseudocode No The paper describes its proposed selection methods using textual descriptions and mathematical formulas, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/nabk89/NAS-with-Proxy-data.
Open Datasets Yes NAS-Bench-1shot1 is used as the primary testing platform to observe the effect of the proxy data on the search performance of three NAS algorithms on CIFAR-10... We conduct thorough experiments across diverse datasets, search spaces, and NAS algorithms. (Datasets mentioned: CIFAR-10, ImageNet, CIFAR-100, SVHN, all standard benchmarks and implicitly publicly available and cited).
Dataset Splits No To construct proxy data of size k, k examples among 50K training examples of CIFAR-10 are selected using selection methods. The selected examples are segregated into two parts: one for updating weight parameters and the other for updating architecture parameters. However, the paper does not specify the exact proportion or number of samples in these two parts.
Hardware Specification Yes For example, executing DARTS [Liu et al., 2019] using our selection method requires only 40 GPU minutes on a single Ge Force RTX 2080ti GPU. Owing to the reduction in search cost, searching on Image Net can be completed in 7.5 GPU hours on a single Tesla V100 GPU when incorporating the proposed selection into DARTS.
Software Dependencies No The paper mentions using 'Pytorch model zoo' but does not specify version numbers for PyTorch or any other software dependencies required to replicate the experiments.
Experiment Setup Yes For a fair comparison, the same hyperparameter settings as those offered in NAS-Bench-1shot1 are used for all the tested NAS algorithms. ... It required 3.8 GPU days, i.e., 11.5 hours with eight V100 GPUs, with a batch size of 1024... PC-DARTS with ... a single V100 GPU with a batch size of 256... the weight decay factors for DARTS and Robust DARTS (L2) during search are set to be 0.0003 and 0.0243, respectively.