Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets

Authors: Hayeon Lee, Eunyoung Hyung, Sung Ju Hwang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to validate Meta D2A framework. First, we compare our model with conventional NAS methods on NAS-Bench-201 search space in Section 4.1. Second, we compare our model with transferable NAS method under a large search space in Section 4.2. Third, we compare our model with other Meta-NAS approaches on few-shot classification tasks in Section 4.3. Finally, we analyze the effectiveness of our framework in Section 4.4.
Researcher Affiliation Collaboration Hayeon Lee1 , Eunyoung Hyung1 , Sung Ju Hwang1,2 KAIST1, AITRICS2, South Korea {hayeon926, eunyoung0301, sjhwang82}@kaist.ac.kr
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. The methods are described in textual form and equations.
Open Source Code Yes Code is available at https://github.com/Hayeon Lee/Meta D2A.
Open Datasets Yes We meta-learn the proposed Meta D2A on subsets of Image Net-1K and neural architectures from the NAS-Bench201 search space. We compile Image Net-1K (Deng et al., 2009) as multiple sub-sets by randomly sampling 20 classes with an average of 26K images for each sub-sets and assign them to each task. We apply our model trained from source database to 6 benchmark datasets such as 1) CIFAR-10 (Krizhevsky et al., 2009), 2) CIFAR-100 (Krizhevsky et al., 2009), 3) MNIST (Le Cun & Cortes, 2010), 4) SVHN (Netzer et al., 2011), 5) Aircraft (Maji et al., 2013), and 6) Oxford IIIT Pets (Parkhi et al., 2012).
Dataset Splits Yes We collect Nτ =1,310/4,230 meta-training tasks for the generator/predictor and 400/400 meta-validation tasks for them, respectively. For CIFAR10 and CIFAR100, we used the training, validation, and test splits from the NAS-Bench-201, and use random validation/test splits for MNIST, SVHN, Aircraft, and Oxford-IIIT Pets by splitting the test set into two subsets of the same size. The validation set is used to update the searching algorithms as a supervision signal and the test set is used to evaluate the performance of the searched architectures.
Hardware Specification Yes Our model is performed with a single Nvidia 2080ti GPU. Search times of Meta D2A for CIFAR-10, CIFAR-100, Aircraft, and Oxford-IIIT Pets are within 57, 195, 77, and 170 GPU seconds on average with a single Nvidia RTX 2080ti GPU respectively
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, specific deep learning framework versions like PyTorch or TensorFlow, or CUDA versions).
Experiment Setup Yes Table 6: Hyperparameter setting of Meta D2A on NAS-Bench-201 Search Space lists values for learning rate, batch size, training epoch, and other model-specific dimensions. Section E.3 also details training strategies such as using embedding features from a pretrained ResNet18 and teacher forcing.