Reproduction results are relatively poor

[Miraclerice]

Hello, I would like to ask you a question. I reproduced the code according to your repo. The dice of the experimental results I ran out are about 4 different from the results of your paper. Is there anything I need to pay attention to? @Junelin2333

[Junelin2333]

I noticed what you said about this. Please check if the results obtained on the test set. If valid set, the result is about 0.86 (Dice Score)

[Miraclerice]

Since running the training script reported an error on line 55 of the training script saying that the task parameter was missing, I added that parameter to Accuracy, not sure if that is correct. Here are my results after training, it doesn't show up very high. Could it be because my dataset is set up incorrectly?

[Miraclerice]

If I'm doing it wrong, I'd appreciate guidance and a solution! @Junelin2333

[Junelin2333]

About error on line 55, it's due to the version of Torchmetrics, use parameters task='binary' is also work, I have update the code in train.py and move some code to engine/wrapper.py.

I'm not sure if you want to ask about the discrepancy between the results in Table 1 and the epoch 53 result in your picture. If yes, I have noted that the result should be obtained on the test set rather than valid set. If not, please contact with me via email.

I have upload a file named evaluate.py, please referring the file and the readme to get the result on the test set.

Best wishes :)

[Miraclerice]

Thank you very much for your answers and repo updates, I hope you are getting better and better in your research journey! @Junelin2333

[Junelin2333]

Don't forget give me a star if my anwsers have helped you! Thanks! Best wishes to your academic, too!