Release of model weights

[LudvigDillen]

Hi,

Would it be possible for you to release the model weights of your models, specifically the ones trained on ISIC2018? I tried running the following
Command: OMP_NUM_THREADS=1 python -u train.py --config config/isic-swin-silv2-inpd-multreg-enp.yaml --test --gpu 0
and got the results:

  test/accuracy         0.4166666567325592
    test/loss            2.399796962738037
  test/loss_ce           1.875800609588623
 test/loss_ce_w          1.875800609588623

test/loss_cfn_entropy 7.643303871154785
test/loss_cfn_entropy_w -0.0007643306162208319
test/loss_multifilt_reg 2.9619715213775635
test/loss_multifilt_reg_w 0.029619721695780754
test/loss_topo_ce 1.9805625677108765
test/loss_topo_ce_w 0.4951406419277191
test/roc_auc 0.45845258235931396
test/topo_accuracy 0.05886243283748627
test/topo_roc_auc 0.52615886926651

I guess this is just from the pre-trained Swin2 backbone, and that I'd have to re-train your model to get up to the performance in the paper.

Thanks for the great work!
Ludde

[Kausta]

Hello,

Thank you for your interest in the work. I uploaded the model weights trained on ISIC2018 to experiments/isic2018/best_val.pt, along with the corresponding config file and the command (which should match config/isic-swin-silv2-inpd-multreg-enp.yaml). The results on this file should match:

	test/accuracy         0.8346560597419739
        test/loss           1.0702425241470337
	test/roc_auc           0.965003490447998
   	test/topo_accuracy       0.8320105671882629
    	test/topo_roc_auc        0.944585382938385

I will upload the remaining weights as well, but it may take me a bit of time. If you need any help or run into any issues, please let me know.
Thanks,

Caner

[LudvigDillen]

Hi there,

Thanks for the weights of TopoSwin-MP. I managed to get these results which largely overlaps with the paper :)

   test/accuracy 0.8320105671882629 (paper has 0.8347, ok)
   test/loss 1.0674318075180054
   test/roc_auc 0.9648759365081787 (paper has 0.9650, ok)
   test/topo_accuracy 0.8339946866035461
   test/topo_roc_auc 0.9449227452278137

Although, I did not understand the command (cmd.txt) you added since it was with the train flag on. Also, I had to debug a bit to realize that I had to add this line to the config file
resume: experiments/isic2018/best_val.pt
But after doing that, everything worked out. Would be very appreciated if you uploaded the weights for "TopoSwin-Betti" and "TopoSwin-SP" for ISIC2018 as well.

I guess "TopoSwin-SP" belongs to the config isic-swin-silv2-inpd-singreg-enp.yaml but which config belong to "TopoSwin-Betti"? And what is the difference between "inpd" and "inponly"?

All the best,
Ludde