marrlab / shapr_torch Goto Github PK

Hello!

Dear colleagues, I'm trying to reproduce results from your paper.
I have downloaded data and executed scripts, assuming that the first one is SHAPR and the second one is SHAPR+topological loss:

python run_train_script.py -p config/red-blood-cell.json
python run_train_script.py -p config/red-blood-cell-2D.json

Training has finished, but numbers on a test set which I see in WanDB are different from numbers in your paper (https://arxiv.org/pdf/2203.01703.pdf, Table 1).
In fact the list of metrics in WanDB (IoU_Error, dice_error, volume_error) are different from the errors presented in the paper (1 − IoU, Volume, Surface area, Surface roughness).
The script "shapr/scripts/evaluation.py" only outputs figures.

Can you please give a hint how to reproduce results?

I have not used wandb before but it seems to be an alternative to traditional Python logging which saves logs to a website? I have not been able to run the code with the references to wandb included (it requires a login and then refuses access to this project). The code seems to work if I comment out any reference to wandb, but rather than submitting a pull request for this, I thought I would submit this Issue in case wandb has been intentionally left in? As far as I can see, there are only references to wandb in the main.py file.

Hi,

In the _settings.py file, we have the default setting "topo_feat_s=False", which means "do no use superlevel features". But in the paper, the authors applied a superlevel set filtration, which means constructing cubical complexes by retaining vertices above a sequence of thresholds.

Could you answer the questions below?

1. which setting for "topo_feat_s" is used
2. from your experience, what is the effect of superlevel and sublevel set filtration?

Best,
Vincent