dncnn "3d"?

[franzpx125]

Hi Allard,

it seems to me that the "msd" approach works in a pure 3D fashion (by setting n_dim=3 when instantiating MSDRegressionModel). Could it be possible to have something similar for "dncnn"? I know that processing becomes prohibitively slow but it would be interesting to test the approach.

Thanks!

[ahendriksen]

Hi Francesco,

That is certainly possible! The simplest way would be to copy the dncnn file and replace the nn.Conv2d calls with nn.Conv3d. For instance, here and on lines 30, and 38.

Beware though, that it is very likely that the method will run out of memory! Counter-intuitively, UNet is slightly more efficient in 3D because a substantial part of the processing takes place on downsampled images/volumes, which require substantially less memory.

[franzpx125]

Ok. I'll give it a go. Should I change also line 36 from nn.BatchNorm2d to nn.BatchNorm3d?

[ahendriksen]

Ok. I'll give it a go. Should I change also line 36 from nn.BatchNorm2d to nn.BatchNorm3d?

Yes, definitely. Good catch!

[franzpx125]

Ok. Thanks! I'm gonna try...