The pretrained weights and data link have expired.

[m-lyon]

The links to the pretrained weights and dataset have expired. Are you able to provide new links?

If you're looking for a place to host datasets/weights for research papers, where the links don't expire, then I would recommend zenodo.org.

[dafhafhajfajfna]

I found that the model I trained was very poor. Have you trained this network?

[m-lyon]

I have yes, I've found the results of the trained model to be on par with those presented within the paper. However, the model i trained used denoised data with patch2self. If your data is not denoised then you will need to train the model yourself, though I can provide a training pipeline to do so.

[dafhafhajfajfna]

Thank you for your answer.
Have you checked the model output of the intermediate process? In order to learn fod_net simply, I only used 10 data for training, and I simply tested 1000 epoch models. I visualized the test results in mrview and found the effect was very poor.Is this because my model hasn't started to converge yet?In addition, I use the preprocessed data of hcp, so the impact of noise on the network should be relatively small

[m-lyon]

I'm not sure what you mean by the intermediate process? Do you mean reconstructing outputs after some amount of epochs but before training has finished?

10 data as in 10 subjects or 10 patches of input? The training regime I used was on the order of around 30 subjects for 100 epochs.

I also used the preprocessed HCP data and a model trained on patch2self denoised data will still perform poorly on non-denoised data as the patch2self denoiser is relatively aggressive.

[dafhafhajfajfna]

I used 10 subjects for1000 epochs and tested the model saved when epochs=75 100 and 1000 respectively, and the results were all poor. I even tested the training data, but the results were still poor. I didn't know what was wrong.

[m-lyon]

Have you calculated the error in AFD or ACC in the reconstructed images at all? This would give a good indication of the level of error you are experiencing.

Even after training, when visualising the resulting FODs, they will often have some noticeable differences compared to the ground truth.

Drop me an email and I can see if i can assist further.

[dafhafhajfajfna]

I have not calculated these indicators. I have sent an individual data I used and the test results to your email

[GuoKefu]

Have you resolved the problem you encountered now？