antoyang / nas-benchmark Goto Github PK

Hello. thank you for the work. i am personally interested in MANAS.But i could not see it implementation
if you add it that will be great

Hi,

I just find there is a mistake in your released DARTS code.

In architect.py, the hessian is computed as follows:
hessian = [(p-n) / 2.*eps for p, n in zip(dalpha_pos, dalpha_neg)]

However, referring to paper and the official release code, it should be computed as follows:
hessian = [(p-n) / (2.*eps) for p, n in zip(dalpha_pos, dalpha_neg)]

But it seems this mistake does not hurt the performance. LOL

hi

I am research the NAS for my master degree.

I am confuse about the paper in Table 2 "nb cells A other datasets " parameter .

This parameter is setting as 8.

Is it reference by the origin NAS paper's setting on imagenet or it's setting by yourself?

Hi @antoyang, Thanks for releasing the benchmark. I am trying to search a cell using DARTS as mentioned in the README. However, once I change the random seed to a different value or sometimes even with the default seed (--seed 2), the search is not working. It becomes idle at the very beginning (training seems to be not working) although it is using GPU memory. Am I missing something here? Can you please comment on this? Thanks.

Thanks for your great work and very powerful opinions!

It is indeed difficult to evaluate how good or really effective a proposed neural architecture search algorithm as it is described. Random sampling in search space is a powerful baseline for comparsion. Not just for image classification, even for more complex tasks such as dense prediction tasks (segmentation, pose estimation ...), NAS evaluation is still frustratingly hard. When applying NAS for larger datasets like MPII and comparing the random sampled architectures from search space (micro and macro) with the first-order gradient-based search method proposed by DARTS, random sampling method still performs surprisingly powerful for human pose estimation architectures searching.

I wonder if you are willing to set up benckmarks for more tasks and datasets in the future?

Hi, I wanted to know while searching for micro cell for cifar100 using ENAS, does the validation and test accuracy improves over time. I am getting extremely low accuracy even if I train for longer period.

in data folder
Sport8_test.csv bocce 22 columns

have a image name test

is it bug ?