DearCaat / MHIM-MIL

[ICCV 2023 Oral] Multiple Instance Learning Framework with Masked Hard Instance Mining for Whole Slide Image Classification
49 stars 3 forks source link

about dsmil code #16

Closed akidway closed 2 months ago

akidway commented 2 months ago

Hi, @DearCaat

Thank you for you nice job.

I have a little question about your code of dsmil. In your code, class FCLayer use an activation function such as gelu or relu, while in official dsmil code, there is no activation function in FCLayer.

Will this nonlinear lead a better result than it actually is.

I wonder if you could share some of your views.

DearCaat commented 2 months ago

Hi, thanks for attention!

In fact, all of the comparison methods (including DSMIL) that appear in the paper are reproduced by code from this repository. So, you can see all the reproduced methods include an activation function in the FCLayer. I think these are fair.

Moreover, u can drop the activation function in the FCLayer. It will have an effect on the final performance, but this effect is present for all methods.

akidway commented 2 months ago

Thank you for reply. Best wishes.

DearCaat commented 2 months ago

Thank you for reply. Best wishes.

Best wishes!