SUMNAS: Supernet with Unbiased Meta-Features for Neural Architecture Search

Authors: Hyeonmin Ha, Ji-Hoon Kim, Semin Park, Byung-Gon Chun

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate SUMNAS with qualitative and quantitative experiments on the CIFAR10 and Image Net datasets.
Researcher Affiliation Collaboration Hyeonmin Ha Seoul National University Ji-Hoon Kim NAVER AI Lab, NAVER Corporation; NAVER CLOVA, NAVER Corporation Semin Park Yonsei University Byung-Gon Chun Seoul National University; Friendli AI
Pseudocode Yes Algorithm 1 Meta-feature training
Open Source Code No No, the paper does not provide a direct link or an explicit statement about the public availability of its source code.
Open Datasets Yes We evaluate SUMNAS on two search spaces NAS-Bench-201 (Dong & Yang, 2020) on CIFAR10 (Krizhevsky et al., 2009) and Mobile Net blocks on Image Net (Russakovsky et al., 2015).
Dataset Splits Yes For the experiments of the NAS-Bench-201 search space (Dong & Yang, 2020) and CIFAR-10 (Krizhevsky et al., 2009), we train the supernets on the entire training set of CIFAR-10. On the test set, the hyperparameters are tuned and the reported Kendall tau and accuracies are measured. We also search for the best architecture on the test set. (...) For the experiment of the Mobile Net-based search space and Image Net (Russakovsky et al., 2015), we tune hyperparameters and search the best architecture using a validation set that includes about 50K examples sampled from the training set.
Hardware Specification No No, the paper does not specify the exact GPU/CPU models, memory, or other specific hardware used for running the experiments.
Software Dependencies No No, the paper does not provide specific software names with version numbers for reproducibility (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes We describe the hyperparameters we used in Appendix D.