Before proto

[wwx13]

Hi!
Sorry, i have too many questions bother you....

class LTA(nn.Module):
  def forward(self, *input):

      elif type == 'proto_adapt':
                  unseen_protos, seen_y = input[:-1]
      
                  device = unseen_protos.device
                  before_protos = torch.zeros(len(self.seen_class_protos), self.emb_size).to(device)
                  for i, x in enumerate(self.seen_class_protos):
                      before_protos[i] = self.seen_class_protos[i]  # matrix S
                  seen_protos = before_protos[seen_y]  # matrix S_i
      
                  protos = torch.cat([seen_protos, unseen_protos], 0)  # matrix R

Why we need before_protos? Can we use torch.select(self.seen_class_protos, seen_y to replace above codes?

[Quareia]

I met some back propagation wrong when indexing/slicing on nn.ParameterList, maybe torch.select can do it~