Hello, I have read your paper and would like to ask you a question： How to underst

Yes exactly. And the model is there <code class="notranslat

Parameterization of displacement fields about naisr HOT 3 CLOSED

WangJYao commented on June 14, 2024

Parameterization of displacement fields

from naisr.

Comments (3)

jiaoyining commented on June 14, 2024

Thanks for asking.
In the code this is the function for getting the displacement:

def get_displacement_per_covariate(self, enc_coords, embedding, covaraite, which_covariate, whether_inv=False, training=True):

        embedding_with_attri = torch.cat((covaraite, embedding), dim=-1)
        if training:
            embedding_with_zeros = torch.cat((torch.randn_like(covaraite)*1e-3, embedding), dim=-1)
        else:
            embedding_with_zeros = torch.cat((torch.zeros_like(covaraite) * 1e-3, embedding), dim=-1)

        arr_disp = self.net_direction[which_covariate](embedding_with_attri, enc_coords) - self.net_direction[which_covariate](embedding_with_zeros, enc_coords)
        if not whether_inv:
            return arr_disp
        else:
            return -arr_disp

self.net_direction() is the f(p, x, z), x is a single covariate for example age.
and arr_disp give the covariate-wise displacement, which is gi(p, x, y, z) = fi(p, x, z) − fi(p, y, z). In this case, y is set to 0 so you get gi=0 when x=0.
During training, we add an extremely small value to 0 as embedding_with_zeros = torch.cat((torch.randn_like(covaraite)*1e-3, embedding), dim=-1) as data augmentation for robustness.

Hope it helps.

from naisr.

WangJYao commented on June 14, 2024

Thank you for your answer !
Is the training code 'train_atlas_3dnaigsr.py' ?

from naisr.

jiaoyining commented on June 14, 2024

Yes exactly. And the model is there naisr/model_naigsr.py

from naisr.

Recommend Projects