DCLP: Neural Architecture Predictor with Curriculum Contrastive Learning

Authors: Shenghe Zheng, Hongzhi Wang, Tianyu Mu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally demonstrate that DCLP has high accuracy and efficiency compared with existing predictors, and shows promising potential to discover superior architectures in various search spaces when combined with search strategies.
Researcher Affiliation Academia Massive Data Computing Lab, Harbin Institute of Technology shenghez.zheng@gmail.com, {wangzh, mutianyu}@hit.edu.cn
Pseudocode Yes Algorithm 1: Training procedure of DCLP
Open Source Code Yes Our code is available at: https://github.com/Zhengsh123/DCLP.
Open Datasets Yes NAS-Bench-101. It is an OON set of 423k CNN (Ying et al. 2019). ... Test accuracy on CIFAR-10 is available within the search space. NAS-Bench-201. This is a NAS benchmark in OOE space (Dong and Yang 2020) with 15,625 structures, providing performance on CIFAR-10, CIFAR-100, and Image Net. DARTS Space. It is an OOE search space (Liu, Simonyan, and Yang 2019)... Language Model. ... DARTS RNN Space: ... PTB as the dataset for the language model task.
Dataset Splits No The paper mentions training and testing but does not explicitly provide details about validation dataset splits or how they are generated, such as percentages, sample counts, or specific methodology for reproduction.
Hardware Specification Yes We use a single RTX3090 as the platform.
Software Dependencies No The paper does not provide specific version numbers for software dependencies such as libraries or frameworks (e.g., PyTorch, TensorFlow, CUDA versions) used for the experiments.
Experiment Setup No We report our setups in Appendix D.