Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Ask for Help about how to achieve fixed sampling and learned sampling without reconstruction? If I just frozee those parameters, gradients can not prapogate. #3

Open
JingYang321 opened this issue Apr 24, 2023 · 15 comments

Comments

@JingYang321
Copy link

No description provided.

@tomer196
Copy link
Owner

Hi,
For fixed sampling without reconstruction, you don't need to run training at all. Just run inference and extract the DWI after subsampling and before the reconstruction network.
For learned sampling, you must run the network and train both the sampling and the reconstruction. After training is done, you can run the inference and again extract the DWI before the reconstruction network.
If you have any more questions feel free to ask.
Tomer

@JingYang321
Copy link
Author

JingYang321 commented Apr 24, 2023 via email

@tomer196
Copy link
Owner

You can't check DWI directly because the downsampled DWI and the full one have a different number of gradient directions (and also different directions).
We also found that it is less meaningful to compare DWI's and evaluate the performance of the end-task is much more powerful.

@JingYang321
Copy link
Author

JingYang321 commented Apr 24, 2023 via email

@JingYang321
Copy link
Author

JingYang321 commented May 5, 2023 via email

@tomer196
Copy link
Owner

tomer196 commented May 8, 2023

I didn't understand your question. Can you explain what you meant?

@JingYang321
Copy link
Author

I actually want to know how to get the corresponding fixed-direction sample file, for example, if I want to get three directions, how do I get file like dir90.mat. Thank you very much.

@tomer196
Copy link
Owner

tomer196 commented May 8, 2023

@JingYang321
Copy link
Author

I checked the code, but I find that these mats are about Learned directions for few acceleration factors. Does this mean the direction mat is learned from the subsample network, rather than using the electrostatic repulsion algorithm?

@tomer196
Copy link
Owner

tomer196 commented May 8, 2023

@JingYang321
Copy link
Author

Oh,I see. SO these mat files in https://github.com/tomer196/Learned_dMRI/tree/master/bvecs from dipy.core.gradients.disperse_charges?

@tomer196
Copy link
Owner

tomer196 commented May 8, 2023

I actually don't remember if this are the fixed or learned directions ;-)

@JingYang321
Copy link
Author

Probably these are learned directions, "We consider the following baselines: learned directions (joint opti-
mization of diffusion directions and the reconstruction network, φ and θ were
initialized randomly), and fixed directions (optimizing the reconstruction net-
work alone). "
from your paper. But I do not know how to train reconstructe network with fixed direction

@tomer196
Copy link
Owner

tomer196 commented May 8, 2023

@JingYang321
Copy link
Author

oh,It is so simple. Thank you very much

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants