Bartosz Ludwiczuk

Hi. I have manage to translate most of the code, which you were using into new version of [code](https://gist.github.com/melgor/7dd78c4e5a7b9322034776535a92c23c) About you second question, I can not figure it out too....

I also faced this problem, but I managed to resolve it. In my case, it was just using `torchelastic` instead of `torch` for starting the jobs as `torchelastic` is now...

Very good catch :) Then only backbone is learned and in classification layer just random weight stay for all training. Author should change this implementation, guided by your explenation.

`sudo ldconfig` should resolve this problem. It just update links to library.

First of all, I advice you to read paper: https://arxiv.org/abs/1801.07698 It contain some summary of L/A-Softmax based idea. Here is very nice plot of ψ functions, based on several ideas....

I also noticed that issue. My current understanding is that in validation batch-size=1 for a single tile (like 512x512) they run 9 predictions (224x224 x9 as input to the network...

I aslo was trying to run the model, but with the success. Here is my path: ``` conda create -n ibm python==3.10 conda activate ibm pip3 install torch torchvision pip...

I managed to run single_mesh.zip package. Based on my investigation it is required to change the python version to a newer one to update setuptools to a new version. I...