search

ajbrock / SMASH

An experimental technique for efficiently exploring neural architectures.
489 stars 57 forks source link

Accuracy only about 27% on CIFAR-100 #5

Open songzhaozhe opened 6 years ago

songzhaozhe commented 6 years ago

Hi,

I like the idea very much, but I am wondering how this work is going since there have been no updates for 4 months.

I ran the 3 instructions twice mentioned in the readme, and achieved only 27.2% and 26.5% val error on CIFAR-100, substantially higher than the paper.

Is this because the default setting is incorrect, or the variance is huge due to randomness?

Thank you!

rainfalj commented 6 years ago

My train on cifar10 just could not reach 4.03%, I use test data and big op