[ICLR'21] DARTS-: Robustly Stepping out of Performance Collapse Without Indicators [openreview]

Authors: Xiangxiang Chu, Xiaoxing Wang, Bo Zhang, Shun Lu, Xiaolin Wei, Junchi Yan

Architecture Search

See scripts/run_darts_minus.sh for searching in S0-S4 (S0 corresponds to S5 in the code) on CIFAR-10 and CIFAR-100.

Evaluation

See scripts/run_darts_minus_fulltrain.sh for the evaluation of CIFAR-100 models in all search spaces and CIFAR-10 models in S4. The rest CIFAR-10 models are evaluated with SGAS code in script/eval.

Hessian Eigenvalue Calculation

During the architecture search, we turn --compute_hessian off by default. Once the search is done, Hessian eigenvalues can be calculated independently from saved checkpoints.

See scripts/start_calc_hessian.sh.

Loss Landscape

See scripts/start_draw_loss_landscape.sh for details.

Recruiting / 招聘

• We are hiring for interns and professionals who are avid in machine learning (especially vision, perception in self-driving, recomendation systems etc.), mailto: zhangbo97(at)meituan.com
• 实习/社招请发简历到zhangbo97(at)meituan.com, 实习JD.

Acknowledgement

Code heavily borrowed from DARTS, RobustDARTS, Loss Landscape, and SGAS.