New approach for weekly meetings?
I thought it would more useful to put my thoughts into text and LaTeX than into powerpoint every week, since it will be easier to keep track of and to put into publications when I write up my research. For update purposes, is this as good as a weekly powerpoint?
Worked on writeup of FMAPICON Rolling Loss
Performance has plateaued on the DAVIS task, although I left it training all week. I put some effort into explaining why the rolling loss helps, along with a more general Writeup of the augmentation approach to equivariance– if I can wrangle this into a paper, then this writing will be useful.
Worked on implementing ICON feature requests
I am struggling with wrangling itk's displacementFieldTransform object to do what I want: the interface between numpy arrays and itk images with vector valued pixels is the pain point: it is loose and performs few checks, and so mistakes cause garbage transforms instead of run-time or compile time errors. Currently, as shown above, I am getting garbage transforms. I am confident that I can power through these errors shortly, but this does make me hesitate to throw the other grad students at this as "the interface to end all interfaces." I should email matt about the problems I am having.
UPDATE: I was casting to long instead of double. The larger issue of poor error checking in this corner of ITK stands, as that sort of type pun should be a runtime error. I may make a pull request. The transforms are still garbage, of course, but believe me they were worse.
UPDATE 2: The stride situation was a disaster. I've fixed it with a hack but definitely need to make a pull request upstream for the way pybuffer interacts with vector valued images.
Up to this point I have gotten displacementFieldTransform working for spacing-1 images, but there is some work to do to make sure it can accomodate all spacings.
We have decided on this approach for accomodating unusual spacings:
resize the two images directly to the neural network input size
process them using the pretrained neural network
postprocess the resulting deformation to be valid for the original spacings
Once I have a correctly initialized itk.displacementFieldTransform then I can use it to warp an image using a very standard itk idiom.
This then makes checkerboard visualization and Dice calculation easy.
The remaining big items on the ToDo list center around training with label losses, atlas losses, and training strategy. We can discuss what do to during today's meeting! I have read zhipeng's paper, he gets excellent registration results
Back to Reports