D-X-Y / AutoDL-Projects

Automated deep learning algorithms implemented in PyTorch.
MIT License
1.56k stars 281 forks source link

The search results of DARTS #58

Closed GongXinyuu closed 4 years ago

GongXinyuu commented 4 years ago

Hi, thanks for your great work! However, I found something strange when I use your code to search.

Describe the bug The search result of DARTS is a combination of skip_connect.

To Reproduce Just run:

CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/DARTS-V1.sh cifar10 1 -1
CUDA_VISIBLE_DEVICES=1 bash ./scripts-search/algos/DARTS-V2.sh cifar10 1 -1

Python version: 3.6, PyTorch version: 1.4.0

Screenshots

image image
D-X-Y commented 4 years ago

That is the expected results and has been discussed in our ICLR 2020 paper (https://openreview.net/forum?id=HJxyZkBKDr). In short, DARTS is easy to converge to a network with many skip connection layers, which has also been found by many other papers.