can we just use update_location_indices to initiate active_sites_map

[978326187]

rpg_asynet/layers/conv_layer_2D.py

Line 67 in e11c0a5

active_sites_map[update_location_indices] = Sites.ACTIVE_SITE.value

rpg_asynet/layers/conv_layer_2D.py

Line 65 in e11c0a5

active_sites_map = torch.squeeze((torch.sum(feature_map**2, dim=-1) != 0).float()).to(self.device)

i think the cases in line 65 is included in line 67.

[MessikommerNico]

Active site are defined as non-zero sites in the input. Later through the network, active site can become zero without being deactivated.
The line at 67 catches the cases, in which active sites became zero through the network update (which happens a lot due to ReLU).

[978326187]

does that mean variable update_location_indices doesn't include non-zero sites in the input ?

[MessikommerNico]

update_location_indices includes non-zero sites. But it only considers the sites, which did change from one timestep to the other. Hence, both lines are necessary to get the active site map of the whole activation map

[978326187]

got it ,thanks very much~