shunlu91/pa-da

[CVPR '23] PA&DA: Jointly Sampling PAth and DAta for Consistent NAS

pa-da's Issues

Questions about probabilistic shape

Respected researchers
I'm sorry to disturb you. I still have a couple of small doubts about PA-DA that I would like to ask you about, thank you very much!

1. Why is an array of [0.2, 0.2, 0.2, 0.2, 0.2, 0.2] used to control the probability of sampling each edge in B-201 instead of an array like shape(5,6)? Is the probability of each edge the same if it is in DARTS space? Can you tell me the principle?
2. If it is a shape(7,15) search space, how should I set up the PA sampling rules?
3. PA-DA reaches Ktau's SOTA, but I still can't run in the 0.713 range, it's roughly around 0.69, and I don't know where I'm not setting the parameters correctly.

Sincerely look forward to your guidance, best wishes to you.

Respected researchers

I am very interested in your research. Could you please share the code related to the darts search space to better reproduce the outcomes in the paper?

where is the data importance sampling in the code？

Hi, I am interested in your work.

Could you please share the data importance sampling code with me?

How were the 64 architectures selected, randomly or carefully?

Dear researcher, I have a few small doubts to ask you.

1. How to choose the 64 architectures, are they random or carefully selected? Were they selected according to a certain ranking arrangement?

2. about the measurement of ktau coefficient, why the Fig. 1(a) is 0.4, while the paper says it is 0.713. Any evaluation rules you can share?

Thank you very much.

nasbench201_dict.npy文件询问。

您好，请问这个nasbench201_dict.npy文件是什么作用，如果我用您的方法在新的数据集上跑，这个文件需要修改什么地方。麻烦您了。

关于计算DA阶段的梯度

作者您好，最近您的代码有部分我理解的不是很清楚；
即关于代码中求解Grad_norm的部分，您代码中是否没有使用求导数的方式得到对应的Samples权重？
with torch.no_grad():
probs = F.softmax(logits, dim=1)
one_hot_targets = F.one_hot(targets, num_classes=args.num_classes)
grad_norm = (torch.norm(probs - one_hot_targets, dim=-1).detach().cpu().numpy())
loss = loss_per_sample.mean() # mean loss to backward ?
还有一个问题就是，如果是回归问题，这里应该如何进行修正呢？
我在IS有看见您的Issue，请问您清楚如何使用pytorh转换您询问的该代码吗？

最近拜读您的代码以及IS代码有这样的一些困惑，期待您的解答！
谢谢！