sgadgil6 / cnslab_fmri Goto Github PK

Hello! Thanks for sharing the source code and it helped me a lot.
I have attempted to reproduce the results by your code while there are some problems.

In the training and testing process in each fold, the model can get expected test results around 0.83 for acc.
I find the model is unstable and train/test loss are still sharply changing in every epoch and even tend to rise. Same problem was found in train/test acc. For example:

      [141001] training loss: 0.310 training batch acc 0.875000
      0.7935779816513762
      test loss = 251.06089854240417
      [142001] training loss: 0.313 training batch acc 0.843750
      0.8119266055045872
      test loss = 251.3142318725586
      [143001] training loss: 0.310 training batch acc 0.875000
      0.7798165137614679
      test loss = 243.6745239496231
      [144001] training loss: 0.313 training batch acc 0.906250
      0.8027522935779816
      test loss = 255.16645431518555
      [145001] training loss: 0.313 training batch acc 0.953125
      0.8027522935779816
      test loss = 243.88370209932327
      [146001] training loss: 0.315 training batch acc 0.875000
      0.8027522935779816
      test loss = 257.25346076488495
      [147001] training loss: 0.314 training batch acc 0.875000
      0.8027522935779816
      test loss = 258.0545355081558
      [148001] training loss: 0.310 training batch acc 0.843750
      0.7889908256880734
      test loss = 253.85232359170914
      [149001] training loss: 0.312 training batch acc 0.796875
      0.7981651376146789
      test loss = 250.11881053447723
      [150001] training loss: 0.310 training batch acc 0.890625
      0.7935779816513762
      test loss = 259.8563167452812
      [151001] training loss: 0.309 training batch acc 0.843750
      0.7935779816513762
      test loss = 257.9779593348503
      [152001] training loss: 0.311 training batch acc 0.890625
      0.8073394495412844
      Best accuracy for window 128 and fold 1 = 0.8394495412844036 at epoch = 27000

Is it appropriate to choose the best test result in some epoch or the final test result after all epoches as the model test acc even if it is still sharply changing?

There is something wrong with your code. There is a missing net.train() and net.eval() in the .py files for st_gcn training. When it is fixed, the results cannot be reproduced.

Hi:
I want to learn how to use final processed BOLD you provided
" https://drive.google.com/file/d/1c4UYvp089KwqAllAHdHMHGjq8y5x882D/view " to run this paper code.

In this run_training_lstm.py,
when running to loss.backward() ,break

if I use retain_graph=True ,then buff will OOM

hello,
thanks for sharing your code.
i'm trying to reproduce results but not getting similar accuracy.

steps i took:

1. download the folder 'hcp_tc_npy_22' from the google drive link that you supplied.
2. run the preprocessing.py script after configuring the folders.
3. run the 'run_training.py' script.

i made minor changes such as configuring 'adj_matrix.npy' isntead of 'adj_matrix_qingyu.npy'. (since this adj_matrix is generated automatically in the preprocessing.py script)

the accuracy doesn't go over 55%. what can possibly explain this?

Thanks for sharing your code. I want to know how to process the download fmri data to the final processed BOLD signal. I am wondering if you could kindly tell me the detail about the preprocessing, such as the atlas, toolbox or processing script. Thanks.

Could you please tell me how the Fig. 3 and Fig.4 in the paper "Spatio-Temporal Graph Convolution for
Functional MRI Analysis" were plotted? Thanks!

Hi, thanks for sharing your excellent work of MICCAI 2020.

I want to re-implement this project, but I don't know how to get the data used in paper.

Will you share or plan to share the data related with this paper?

Thank you!