Differentiable Neural Architecture Search in Equivalent Space with Exploration Enhancement

Authors: Miao Zhang, Huiqi Li, Shirui Pan, Xiaojun Chang, Zongyuan Ge, Steven Su

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compared the proposed approach with different One Shot NAS baselines on the NAS benchmark dataset NAS-Bench-201 [13], and extensive experimental results illustrate the effectiveness of our method, which outperforms all baselines on this dataset. and 4 Experiments The high computational cost of evaluating architectures is the major obstacle of analyzing and reproducing One-Shot NAS methods, and it is hard to reproduce current NAS methods under the same experimental setting for a fair comparison. In this section, we adopt the NAS-Bench-201 [13] as the benchmark dataset to analyze our E2NAS.
Researcher Affiliation Collaboration Miao Zhang1,2,3, Huiqi Li1 , Shirui Pan3 , Xiaojun Chang3, Zongyuan Ge3,4, Steven Su2 1Beijing Institute of Technology 2University of Technology Sydney 3Monash University 4Airdoc Research Australia
Pseudocode Yes Algorithm 1 E2NAS 1: Input: Trained encoder E and decoder D, training dataset Dtrain and validation dataset Dvalid. 2: Initial architecture archive A = ; 3: Randomly initialize architecture parameter αθ and supernet weights WA(α); 4: while not done do 5: Sample batch of Dtrain, decode αθ to get α based on D, get the complementary architecture αc, and update the supernet weights WA(α) based on Eq. (7), and add architecture α into A; 6: Sample batch of Dvalid, and update αθ based on Eq. (3); 7: end while 8: Decode αθ to obtain the best α based on the trained decoder D. 9: Retrain α and get the best performance on the test dataset Dtest. 10: Return: architecture α with best performance.
Open Source Code No The paper does not contain any explicit statement about releasing source code or a link to a repository for the described methodology.
Open Datasets Yes In this section, we adopt the NAS-Bench-201 [13] as the benchmark dataset to analyze our E2NAS. and The search space in NAS-Bench-201 contains four nodes with five associated operations, resulting in 15625 cell candidates.
Dataset Splits Yes In this section, we adopt the NAS-Bench-201 [13] as the benchmark dataset to analyze our E2NAS. and Method CIFAR-10 CIFAR-100 Image Net-16-120 Valid(%) Test(%) Valid(%) Test(%) Valid(%) Test(%)
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as library names or exact programming language versions.
Experiment Setup Yes The hyperparameters of E2NAS are set as ε=0.5 and γ=Sigγ(10) in this experiment. and Generally, there are only two additional parameters that need to be specified, ε for the supernet training and λ for architecture parameter learning, in our E2NAS. and We set eight different settings for γ, including four static settings and four Sigmoid-type settings, as shown in the third block of Table 2.