Method BatchGraphDeepFMPredict

src/tools/model_server.cc:301–346

}

bool ModelServer::BatchGraphDeepFMPredict(
    const std::vector<features_t>& batch_features,
    const std::vector<features_t>& batch_users,
    std::vector<float>* batch_prob) const {
  DXCHECK(target_type_ == 1);
  if (batch_features.empty()) {
    return false;
  }

  if (!graph_ || !model_) {
    return false;
  }

  OpContext op_context;
  op_context.Init(graph_.get(), model_->mutable_param());
  if (!op_context.InitOp({target_name_}, -1)) {
    return false;
  }

  Instance* inst = op_context.mutable_inst();
  auto& X = inst->insert<csr_t>(deepx_core::X_NAME);
  for (const auto& features : batch_features) {
    EmplaceRow(features, &X);
  }

  auto& User = inst->insert<csr_t>(instance_name::X_USER_NODE_NAME);
  for (const auto& users : batch_users) {
    EmplaceRow(users, &User);
  }
  DXASSERT(X.row() == User.row());

  inst->set_batch(X.row());

  op_context.InitPredict();
  op_context.Predict();
  const auto& P = op_context.hidden().get<tsr_t>(target_name_);
  DXASSERT(P.is_rank(2));
  DXASSERT(P.same_shape(X.row(), 1));
  batch_prob->resize(X.row());
  const float_t* _P = P.data();
  for (int i = 0; i < X.row(); ++i) {
    (*batch_prob)[i] = (float)*_P;
    ++_P;
  }
  return true;
}

static void DeleteOpContext(OpContext* op_context) noexcept {
  delete op_context;