About search cost: if you use a Titan X with 12GB memory, the batch size might be reduced to 64.
It is strange that the search cost on 1080-ti is less than 12 hours but about 24 hours on Titan X(p).
This may due to different architectures of those two GPUs.
About arch: in our experiments, this happened with a very small chance, especially with small batch size.
This is because our method is built on top of DARTS and sometimes we meet some unfriendly local minima.
by using default parameters![图片](https://user-images.githubusercontent.com/19514053/64686683-a1cc3b80-d4bb-11e9-8913-98e80a96a469.png)