Skip to content

Remove useless mask in InteractionGNN2 model

Sylvain Caillou requested to merge sc_remove_useless_mask_in_ignn2 into dev

This MR remove the useless following lines from the beginning of the forward method of the InteractionGNN2 model :

# Same features on the 3 channels in the STRIP ENDCAP TODO: Process it in previous stage

mask = torch.logical_or(batch.hit_region == 2, batch.hit_region == 6).reshape(-1)

x[mask] = torch.cat([x[mask, 0:4], x[mask, 0:4], x[mask, 0:4]], dim=1)

# print(x[:, 8:12])

This was a hard-coded workaround mask introduced to temporarily ensure compatibility with the CHEP23 results, where cluster information was used alongside space point information, but only in the strip barrel. However, it has since been shown that incorporating this information from the entire strip regions, including the endcaps, achieves the same or better GNN performance. As a result, these lines are no longer needed.

Merge request reports