However, the number of extraction levels should be set as extract_levels=[3, 2, 1, 0] in order for the average over feauture maps to be takes over all possible resolutions. With extract_levels=[3]the average is only taken over the coarsest (deepest) feature map, meaning a very coarse displacement field is predicted by the network, impacting performance negatively.
I would further recommend that out_kernel_initializer is set to "kaiming_uniform", as this is the default value in the LocalNet constructor (not "zeros"), and we have seen instabilities during training when using "zeros". An additional proposal would be to use an additional layer in the network, and adjust num_channel_initial, accordingly.
The LocalNet in the 3D registration tutorial is currently instantiated as:
However, the number of extraction levels should be set as
extract_levels=[3, 2, 1, 0]
in order for the average over feauture maps to be takes over all possible resolutions. Withextract_levels=[3]
the average is only taken over the coarsest (deepest) feature map, meaning a very coarse displacement field is predicted by the network, impacting performance negatively.I would further recommend that
out_kernel_initializer
is set to"kaiming_uniform"
, as this is the default value in theLocalNet
constructor (not"zeros"
), and we have seen instabilities during training when using"zeros"
. An additional proposal would be to use an additional layer in the network, and adjustnum_channel_initial
, accordingly.All in all, the instantiation would become:
Or simplified (using defaults):