NAS-Bench-101: Towards Reproducible Neural Architecture Search

Authors: Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, Frank Hutter

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We trained and evaluated all of these architectures multiple times on CIFAR-10 and compiled the results into a large dataset of over 5 million trained models. This allows researchers to evaluate the quality of a diverse range of models in milliseconds by querying the precomputed dataset. We demonstrate its utility by analyzing the dataset as a whole and by benchmarking a range of architecture optimization algorithms.
Researcher Affiliation Collaboration 1Google Brain, Mountain View, California, USA 2Department of Computer Science, University of Freiburg, Germany.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Moreover, the data, search space, and training code is fully public 1, to foster reproducibility in the NAS community. 1 Data and code for NAS-Bench-101 available at https:// github.com/google-research/nasbench. The code, implemented in Tensor Flow, along with all chosen hyperparameters, is publicly available at https: //github.com/google-research/nasbench.
Open Datasets Yes We trained and evaluated all of these architectures multiple times on CIFAR-10 and compiled the results into a large dataset of over 5 million trained models. [...] We trained and evaluated a large number of different convolutional neural network (CNN) architectures on CIFAR-10 (Krizhevsky & Hinton, 2009)
Dataset Splits Yes All models are trained and evaluated on CIFAR-10 (40k training examples, 10k validation examples, 10k testing examples), using standard data augmentation techniques (He et al., 2016).
Hardware Specification Yes All models were trained on the TPU v2 accelerator.
Software Dependencies No The paper mentions "Tensor Flow" but does not provide a specific version number, nor does it list versions for any other software or libraries.
Experiment Setup Yes We utilize a single, fixed set of hyperparameters for all NAS-Bench-101 models. [...] All models are trained and evaluated on CIFAR-10 (40k training examples, 10k validation examples, 10k testing examples), using standard data augmentation techniques (He et al., 2016). The learning rate is annealed via cosine decay (Loshchilov & Hutter, 2017) to 0 in order to reduce the variance between multiple independent training runs. Training is performed via RMSProp (Tieleman & Hinton, 2012) on the cross-entropy loss with L2 weight decay. [...] we trained all our architectures with four increasing epoch budgets: Estop {Emax/33, Emax/32, Emax/3, Emax} = {4, 12, 36, 108} epochs.