search

yenchanghsu / Generalized-ODIN

6 stars 0 forks source link

dropout on the input ? #5

Open sesmae opened 3 years ago

sesmae commented 3 years ago

Hi,

I have read your paper. The divisor/dividend approach is very interesting. I have a question regarding the dropout regularization that you mentioned in the paper. It says in the paper that 'apply dropout at the inputs for the dividend/divisor structure.'. By this you mean the dropout is applied to the input x itself or to the f_p(x)? btw is there any implementation of your work that I can refer to?

thanks

yenchanghsu commented 3 years ago

The dropout is between the f^p(x) and h(x)/g(x). So you could say it is on the outputs of f*p(x) or the inputs of the dividend/divisor structure. I do not get permission to share the code. Please feel free to send me an email, and I will try my best to help you.

sesmae commented 3 years ago

Thank you for your response, I think the ood datasets that are used for evaluation are far ood to cifar10 and cifar100. Have you tried the algorithm on near ood? like for cifar10 vs cifar100 case? In case I want to use G-ODIN for near ood setting, do you think the search range for the noise magnitude should be changed? also regarding the backbone network for near ood, do you think densenet is sufficient or resnet is a better choice?

yenchanghsu commented 3 years ago

Table 4 of the paper has a setting close to your description. The rows in the table are roughly arranged with the near-to-far order.