Small question about DeepAtlas

[mikami520]

I am now working on the DeepAtlas tutorial and have a quick question about the warp and warp_nearest in the training. Since in the tutorial, you use the alternate training mode to save memory, which means one network will be frozen while the other one is active. In the segmentation network trail, the comment (shown below) said that we could use differentiable warp() instead of warp_nearest() to achieve the joint training, but this case, the registration network is frozen, how could we do the backpropagation and gradient descent for registration network based on this learnable warp()?

In my opinion, we should not freeze the registration network this time if we want to use warp() to train the registration jointly.

Hope someone could help me and thank you in advanced!