BRP-NAS: Prediction-based NAS using GCNs

Authors: Lukasz Dudziak, Thomas Chau, Mohamed Abdelfattah, Royson Lee, Hyeji Kim, Nicholas Lane

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show that our proposed method outperforms all prior methods on NAS-Bench-101 and NASBench-201, and that our predictor can consistently learn to extract useful features from the DARTS search space, improving upon the second-order baseline.
Researcher Affiliation Collaboration 1 Samsung AI Center, Cambridge, UK 2 University of Cambridge, UK 3 University of Texas at Austin, US
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes We make Lat Bench and the source code of Eagle available publicly1 https://github.com/thomasccp/eagle
Open Datasets Yes Throughout the paper, we focus on NAS-Bench-201 dataset which includes 15,625 models. We also compare to prior work on NAS-Bench-101 [32] and DARTS search space [33]. We also introduce Lat Bench, the first large-scale latency measurement dataset for multi-objective NAS.
Dataset Splits Yes All predictors are trained for 100 times, each time using a randomly sampled set of 900 models from the NAS-Bench-201 dataset. 100 random models are used for validation and the remaining 14k models are used for testing.
Hardware Specification Yes We use desktop CPU, desktop GPU and embedded GPU to refer to the devices used in our analysis, with device details described in Section 6. (i) Desktop CPU Intel Core i7-7820X, (ii) Desktop GPU NVIDIA GTX 1080 Ti, (iii) Embedded GPU NVIDIA Jetson Nano, (iv) Embedded TPU Google Edge TPU, (v) Mobile GPU Qualcomm Adreno 612 GPU, (vi) Mobile DSP Qualcomm Hexagon 690 DSP.
Software Dependencies No The paper mentions software components but does not provide specific version numbers for reproducibility (e.g., 'GCN can handle any set of neural network models.').
Experiment Setup Yes Our GCN predictor has 4 layers of GCNs, with 600 hidden units in each layer, followed by a fully connected layer that generates a scalar prediction of the latency. All predictors are trained for 100 times, each time using a randomly sampled set of 900 models from the NAS-Bench-201 dataset. 100 random models are used for validation and the remaining 14k models are used for testing. We use α = 0.5 for all our experiments.