Thanos-DB
commented
1 year ago Hi kingo233,
at the time of writing the paper and even after the paper got accepted there were no available ground truths for the test set. So, what we did was train on the set with available ground truths, predict using the official test set, generate the required files and upload them on the evaluation server. We have never tested the official test set on our own because that was not possible at the time. After submitting the required files the evaluation server would calculate everything and return some statistics (in our paper there is a link with those). So, at first we split the 100 patients as seen in the paper resulting in train, validation and test set with ground truths as other papers do. Then to show that we did not overfit on this test set we predict using the official test set (remember at the time there were no ground truths) and report those too. You can see in the paper that your 71.7% is too low. Maybe you mixed the two test sets somehow and this is why it is this low? I am sure the evaluation platform would not have ~20% difference. Concerning your PyTorch implementation did you change the multiheadattention as discussed in your previous issue? I suspect this is missing to get the remaining 2-3%.
BR Thanos
kingo233
Susu0812
之后，使用这个拆分#6将训练文件夹分成3个部分。我复制了您的结果： 但是当我在患者编号从 101 到 150 的测试文件夹上测试模型时，结果很低： 
ACDC dataset now have two folders: training and testing
After using this split https://github.com/Thanos-DB/FullyConvolutionalTransformer/issues/6 to divide the training folder into 3 parts .I reproduced your result:
But when I tested the model on test folder which patient number from 101 to 150.the result is low:
Did all the results on the ACDC leaderboard test on the testing folder or on the training folder, like in your paper? What's more,in my this repo https://github.com/kingo233/FCT-Pytorch. I trained using the training folder and get dice 90 on testing folder.But I can't get improvement any more...is this the upper limit of FCT?