Bridge the Gap Between Architecture Spaces via A Cross-Domain Predictor

Authors: Yuqiao Liu, Yehui Tang, Zeqiong Lv, Yunhe Wang, Yanan Sun

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we will empirically investigate the effectiveness of the proposed method, including the experiments conducted on Image Net [12] and CIFAR-10 [30], and the ablation studies.
Researcher Affiliation Collaboration Yuqiao Liu 1, Yehui Tang 2,3, Zeqiong Lv1, Yunhe Wang3, Yanan Sun 1 1College of Computer Science, Sichuan University 2School of Artificial Intelligence, Peking University 3Huawei Noah s Ark Lab
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The source code will be available 3. 3https://github.com/lyq998/CDP (Pytorch) https://gitee.com/mindspore/models/tree/master/research/cv/CDP (Mind Spore)
Open Datasets Yes Our experiments are conducted with Py Torch [41] and Mind Spore [25]. Architecture spaces for CDP: NAS-Bench-101 and NAS-Bench-201 are chosen as the source spaces and DARTS is the target space. (...) Image Net [12] and CIFAR-10 [30]
Dataset Splits No The paper mentions 'training labels' and a 'validation dataset' in theoretical discussions (Section 3.4), but does not provide specific dataset split information (percentages, sample counts, or defined standard splits) for the experiments conducted in the paper.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper states 'Our experiments are conducted with Py Torch [41] and Mind Spore [25],' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Search strategy: Following the convention in neural predictor [55], a large set of architectures in the target search space are randomly sampled and their performance is predicted. In addition, according to the previous work on DARTS search space [5], we limit the max number of skip connections to 2. (...) Moreover, we train these shallow architectures for 50 epochs to further save cost.