Regarding SpatialEmbed project

[adi1999]

Hey @charlotte12l , hope you're doing good! I wasn't able to open an issue on the SpatialEmbeddings repo that you've modified for multiclass exp, So wanted to ask whether you were able to obtain desired performance with you implementation which you've provided. And how did you were able to generate foreground weight that you're using in your code in the loss function.
Thank you!!

[charlotte12l]

Hi:

Thanks for your interest! I can't reproduce the paper's performance with my codes so, in the end, I gave up.

http://www.robesafe.uah.es/personal/eduardo.romera/pdfs/Romera17iv.pdf
weight_C = 1/log(1.1+p_C) , p_C = the prob of class_C

I think I generate the foreground weight using this equation. I did that project a long time ago and I've forgotten most of the things, I'm really sorry. Maybe you can follow davyneven/SpatialEmbeddings#5 . Best luck!

[jyang68qq]

@adi1999 Hi, did you successfully get the multi-class setting solved?

[adi1999]

@charlotte12l Thank you for your response!!

[adi1999]

@adi1999 Hi, did you successfully get the multi-class setting solved?

hey @jyang68qq , unfortunately not, I was experimenting to improve the performance by changing architecture and loss function, in which I was successful, but only for car instance, Was trying to evaluate performance on multiclass, but unfortunately it's to working, I'll update in case I'm able to do so!!