| 46 | {} |
| 47 | |
| 48 | std::unique_ptr<model> |
| 49 | CheckpointBinary::get_partner_model(model const& m, |
| 50 | El::Int partner_trainer, |
| 51 | size_t /*step*/) |
| 52 | { |
| 53 | auto const& comm = *m.get_comm(); |
| 54 | |
| 55 | // Start by copying this model, then do the exchange. |
| 56 | auto partner_model_ptr = std::make_unique<model>(m); |
| 57 | model& partner_model = *partner_model_ptr; |
| 58 | |
| 59 | // Keep track of weights that shouldn't be exchanged |
| 60 | std::unordered_map<std::string, std::unique_ptr<weights>> restore_weights; |
| 61 | auto const& weights_names = this->weights_names(); |
| 62 | if (!weights_names.empty()) { |
| 63 | for (auto w : partner_model.get_weights()) { |
| 64 | if (weights_names.find(w->get_name()) == weights_names.cend()) { |
| 65 | using TensorDataType = DataType; |
| 66 | using WeightsType = data_type_weights<TensorDataType>; |
| 67 | restore_weights[w->get_name()] = |
| 68 | std::make_unique<WeightsType>(dynamic_cast<WeightsType&>(*w)); |
| 69 | } |
| 70 | } |
| 71 | } |
| 72 | exchange(comm, partner_model, partner_trainer); |
| 73 | restore_model_weights(partner_model, restore_weights); |
| 74 | |
| 75 | return partner_model_ptr; |
| 76 | } |
| 77 | |
| 78 | } // namespace ltfb |
| 79 | } // namespace lbann |
nothing calls this directly
no test coverage detected