Dan Luo

Thanks for your quick response. I use the FGSM and eps =0.3. I also follow the tutorial of cleverhans, but I also see the same issue. Though this issue is...

Much appreciation. I even doubt whether I am the only one who really encounters with this issue. I honestly appreciate your help if you would like to investigate this issue...

> maybe you trained it for very long time? Yes. My model on MNIST is trained with 100 epochs, and also if I remember correctly, the tutorial on Cleverhans train...

> I'll dig a bit more on this and see if there's a way to add some functionalities in advertorch for this. I think a quick fixup is to shift...

Hi @gwding, I think we can inherit the ideas from cleverhens that checks whether the prediction is from logits, or probability after softmax. I assume it is a similar case...

Referring to https://github.com/RUCAIBox/RecBole/issues/830, here is my solution. Please be free to let me know if something is missing: The function `full_sort_predict` will generate the top_k item list for the testing...

I think you may refer to "hybrid retrieval model", which combines both sparse retrieval, such as bm25, and dense retrieval, as the feature output from this repo.