conversion of cudipy.align.imwarp.DiffeomorphicMap to dipy.align.imwarp.DiffeomorphicMap

[rajnraj100]

Hi...Can we transform cudipy.align.imwarp.DiffeomorphicMap to dipy.align.imwarp.DiffeomorphicMap?

[skoudoro]

Hi @rajnraj100,

it seems to be exactly the same so yes you can transform cudipy.align.imwarp.DiffeomorphicMap to dipy.align.imwarp.DiffeomorphicMap

[rajnraj100]

If I use Diffeomorphic registration part in CPU using dipy, I would get:

from dipy.align.metrics import CCMetric
from dipy.align.imwarp import SymmetricDiffeomorphicRegistration
metric = CCMetric(3)
level_iters = [10, 10, 5]
pre_align = 4x4 numpy array (numpy.ndarray)
sdr = SymmetricDiffeomorphicRegistration(metric, level_iters)
cpu_mapping = sdr.optimize(static, moving, static_affine, moving_affine, pre_align)
type(cpu_mapping )
dipy.align.imwarp.DiffeomorphicMap

But, I have done my Diffeomorphic registration part in GPU using cudipy for speeding up my code. So, I am getting:

from cudipy.align.metrics import CCMetric
from cudipy.align.imwarp import SymmetricDiffeomorphicRegistration
metric = CCMetric(3)
level_iters = [10000, 1000, 100]
sdr = SymmetricDiffeomorphicRegistration(metric, level_iters)
gpu_mapping = sdr.optimize( static_arr, moving_arr, static_affine, moving_affine, pre_align )
type(gpu_mapping )
cudipy.align.imwarp.DiffeomorphicMap

Here, I stored mapping for all my subjects' data. I will apply mapping.transorm(.) or mapping.transform_inverse(.) later as per my need.
Now, for some reason, I need to switch to cpu for rest of my code but for that I need to convert this cudipy.align.imwarp.DiffeomorphicMap to dipy.align.imwarp.DiffeomorphicMap so that I will not have to use gpu anymore.
Here, I am stuck as I do not know how to change it from cudipy (GPU) to dipy (CPU):

cpu_mapping = some_cmd(gpu_mapping )

If it is possible , kindly help me with the command/ code. Thank you.

[pjsjongsung]

You could transfer all the parameters of the DiffeomorphicMap that are located in the gpu(cudipy class) and copy them to the cpu(dipy class). I know it is not trivial, but if you can copy all the arrays that were calculated through optimize, it should be serve as a workaround for now.

[rajnraj100]

This will not serve the purpose. Anyways, thank you for responding.

[skoudoro]

this will not serve the purpose. Anyways, thank you for responding.

not sure why? Can you explain why it will not serve the purpose. I believe you just need to initialize the DiffeomorphicMap and then update the DiffeomorphicMap.forward and DiffeomorphicMap.backward field. I believe nothing more is needed between both implementation.

when applying the transformation backwardor forward will be used.

[arokem]

I'm not so sure that I understand what the word "transform" means here. Do you mean whether you could use them interchangeably? I think the answer to that is yes. If the question is whether we should replace our current implementation in DIPY with the one implemented in cuDIPY, I think that I would be against that, at least for the time being. I think that we want to keep the GPU-accelerated functionality separate for now, because of the complexities of installing and running these implementations.

[rajnraj100]

If I use Diffeomorphic registration part in CPU using dipy, I would get:

from dipy.align.metrics import CCMetric
from dipy.align.imwarp import SymmetricDiffeomorphicRegistration
metric = CCMetric(3)
level_iters = [10, 10, 5]
pre_align = 4x4 numpy array (numpy.ndarray)
sdr = SymmetricDiffeomorphicRegistration(metric, level_iters)
cpu_mapping = sdr.optimize(static, moving, static_affine, moving_affine, pre_align)
type(cpu_mapping )
dipy.align.imwarp.DiffeomorphicMap

But, I have done my Diffeomorphic registration part in GPU using cudipy for speeding up my code. So, I am getting:

from cudipy.align.metrics import CCMetric
from cudipy.align.imwarp import SymmetricDiffeomorphicRegistration
metric = CCMetric(3)
level_iters = [10000, 1000, 100]
sdr = SymmetricDiffeomorphicRegistration(metric, level_iters)
gpu_mapping = sdr.optimize( static_arr, moving_arr, static_affine, moving_affine, pre_align )
type(gpu_mapping )
cudipy.align.imwarp.DiffeomorphicMap

Here, I stored mapping for all my subjects' data. I will apply mapping.transorm(.) or mapping.transform_inverse(.) later as per my need.
Now, for some reason, I need to switch to cpu for rest of my code but for that I need to convert this cudipy.align.imwarp.DiffeomorphicMap to dipy.align.imwarp.DiffeomorphicMap so that I will not have to use gpu anymore.
Here, I am stuck as I do not know how to change it from cudipy (GPU) to dipy (CPU):

cpu_mapping = some_cmd(gpu_mapping )

If it is possible , kindly help me with the command/ code. Thank you.