Could you share the serving code? about inttower

[archersama]

For company privacy, we can't provide serving code. We don't use faiss to compute maxsim score, we have our own container to save embedding,and maxsim score is computed out of container. To compute maxsim score in a parallel way, just add FE-layer in user every layer and item tower last layer and compute maxsim score. In our code ,the multi head maxsim score is computed in a parallel way

self.Fe_linears = nn.ModuleList(
          [nn.Linear(hidden_units[i], self.field_dim*self.user_head) for i in range(len(hidden_units) - 1)])

def fe_score(user_rep, item_rep, user_fea_col, item_fea_col, user_embedding_dim, item_embedding_dim):
  score = []
  for i in range(len(user_embedding_dim)):
    user_temp = torch.reshape(user_rep[i], (-1, user_fea_col, user_embedding_dim[i]))
    item_temp = torch.reshape(item_rep[-1], (-1, item_fea_col, item_embedding_dim[i]))
    score.append((user_temp @ item_temp.permute(0, 2, 1)).max(2).values.sum(1))
  score = torch.stack(score).transpose(1, 0)
  return torch.sum(score, 1)

[thinkwee]

got it. thx for your reply~