DARTS+: Improved Differentiable Architecture Search with Early Stopping
Info
Conference: Arxiv
Cites: 9
Github Stars: -- (paper said it's only need to change few lines of DARTS)
Github Solved/Issue: --
Author:
Main Idea
Darts + Early Stop
After training for a while, Darts will generate too many skip connections which leads to bad performance.
They stop seaching when a sub-block has 2 more skip-connections
Comments
Actually I think this is the same problem as Fair-Nas mentioned before. When early layers learned to discriminate the features, skip connection in late layer will have nice performance than not well trained conv block.
GreedyNAS: Towards Fast One-Shot NAS with Greedy Supernet
Info
Conference: CVPR 2020
Cites: 0
Github Stars: --
Github Solved/Issue: --
Author:
Main Idea
Uniform sampling in NAS is inefficient.
First use uniform sampling train for 10K iterations, then use a greedy method that sample m path and keep the best(best on val) k path in pool, sample path in pool with more probability.