some flaw in "build_template.py"

[tanzhengtang]

Thanks for developing a tool with fast registration speed，some problems were found in the process of using build_template.py.
Is there any error in the compose_warp_inverse parameter on line 308 of syn.py? If multi-scale scale is used, it will cause the returned "rev_inv_warp_field" to not match the scale of compose_warp on line 315.
Some errors also occurred in the use of functions related to the "shape_avg" function, such as the "shape_averaging_invwarp" function in "warputils.py". In line 318, "inverse_warp = reg.warp.get_inverse_warp(n_iters=20)", "reg.warp.get_inverse_warp" is not implemented.
In line 222 of "greedy.py", setting the size in "F.affine_grid" function to "fixed_images.shape" will also result in an error.

The modified code picture is attached below. I don’t really understand the relevant implementation principles. I just made a simple attempt to modify it and ran it through, so I don’t know whether the modification is correct or not.

In addition, using the mi type loss function will overflow the gpu memory(48GB) at an image size of (420,320,520), while other types of loss functions will not. Can it be optimized?

[rohitrango]

Hey there,

Thank you for using FireANTs!

I did a major refactor recently, some of the functions were botched. I haven't added a testcase to verify if template building works, will do it this weekend and upload the updated code. Hopefully the updated code should work this upcoming Monday.

re: mi loss function, I have a very efficient implementation (private for now) but that will be out soon in the next two weeks. Please stay tuned!

[rohitrango]

btw, the image size ((420,320,520)) is very interesting - around 10 times larger than standard in-vivo "toy" datasets like OASIS. What data or applications are you working on?

I'm curating a "Hall of Fame" soon for challenging real-life applications, and would be interested to showcase your application. 🎉

[tanzhengtang]

Hey there,

Thank you for using FireANTs!

I did a major refactor recently, some of the functions were botched. I haven't added a testcase to verify if template building works, will do it this weekend and upload the updated code. Hopefully the updated code should work this upcoming Monday.

re: mi loss function, I have a very efficient implementation (private for now) but that will be out soon in the next two weeks. Please stay tuned!

Sorry for delayed reply, looking forward to the modified code. Another question is why the consumption of using cc in ants is higher than that of using mi, while the situation is the opposite when using fireants?

[rohitrango]

because ants and fireants have completely different computation mechanisms (itk filters and pytorch do things differently under the hood)

however CPU-to-GPU comparisons are not apples-to-apples comparison so i wouldnt be too quick to compare ants and fireants on that front

[tanzhengtang]

Hi, for the latest code, new problem were found in the process of using build_template.py.

[rohitrango]

thank you for the very fast bug report! I have patched it - works on my end. can you try again?

[tanzhengtang]

thank you for the very fast bug report! I have patched it - works on my end. can you try again?

There has a another question. All images has same size, spacing and orientation. Only input one image, the script just work.

[rohitrango]

why are you trying to create an atlas with just one image? 🙂

[tanzhengtang]

why are you trying to create an atlas with just one image? 🙂

My statement may not be clear. When the number of subjects is larger than 1, it cause this error.

[rohitrango]

oh my bad! I had fixed it but didnt push it. Pushed it - it was a minor change (just use the BatchedImages object instead of the tensor)

[tanzhengtang]

The average template function could work, but there has new error....... And registration speed seems to be slower than before.

[rohitrango]

fixed.

I just realized that some of the newer fused operations are somewhat brittle when inputs are broadcasted, etc. I have made a note of it, and will fix it in a future iteration.

do u mind writing some testcases for the template building code, since you have helped quite a bit with helping me find all the edge cases?

[rohitrango]

Hello there! Does the code seem to work now?

[tanzhengtang]

Hello there! Does the code seem to work now?

Sorry for the late reply, we were on holiday so ... do not to test it. The vacation ends next week.

[tanzhengtang]

Hi, another error is about the video memory overflow. If multiple samples are iterated, there are something occupying the video memory and its space is constantly increasing when new sample enters the process. I don't know if this description is appropriate, but it can be guaranteed that the space required for each iteration and every sample is completely sufficient. This problem also occurred in earlier versions, at that time I modified some source code to solve the problem, but deleted all that code because you released the latest code.

This occurs when the third sample enters the process.