code/scripts/bert_joint.py [187:242]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                    ls = icsl_loss_function(intent_pred, slot_pred, intent_label, slot_label, valid_length - 2).mean()

                    if accumulate:
                        ls = ls / accumulate
                    ls.backward()
                    step_loss += ls.asscalar()

            # update
            if not accumulate or (batch_id + 1) % accumulate == 0:
                trainer.allreduce_grads()
                nlp.utils.clip_grad_global_norm(params, 1)
                trainer.update(1, ignore_stale_grad=True)

            if (batch_id + 1) % log_interval == 0:
                toc = time.time()
                # update metrics
                ic_metric.update([intent_label], [intent_pred])
                sl_metric.update(*process_seq_labels(slot_label, slot_pred, ignore_id=pad_label_id))
                log.info('Epoch: {}, Batch: {}/{}, speed: {:.2f} samples/s, lr={:.7f}, loss={:.4f}, intent acc={:.3f}, slot acc={:.3f}'
                         .format(epoch_id,
                                 batch_id,
                                 len(train_dataloader),
                                 log_num / (toc - tic),
                                 trainer.learning_rate,
                                 step_loss / log_interval,
                                 ic_metric.get()[1],
                                 sl_metric.get()[1]))
                tic = time.time()
                step_loss = 0
                log_num = 0

        mx.nd.waitall()
        epoch_toc = time.time()
        log.info('Time cost: {:.2f} s, Speed: {:.2f} samples/s'
                 .format(epoch_toc - epoch_tic, total_num/(epoch_toc - epoch_tic)))
        # evaluate on development set
        log.info('Evaluate on development set:')
        intent_acc, slot_f1 = evaluate(model=model, eval_input=dev_input)
        if slot_f1 > best_score[1]:
            best_score = (intent_acc, slot_f1)
            model.save_parameters(os.path.join(model_dir, model_name + '.params'))


def evaluate(model=None, model_name='', eval_input=''):
    """Evaluate the model on validation dataset.
    """
    ## Load model
    bert, vocabulary = nlp.model.get_model('bert_12_768_12',
                                           dataset_name='wiki_multilingual_uncased',
                                           pretrained=True,
                                           ctx=ctx,
                                           use_pooler=False,
                                           use_decoder=False,
                                           use_classifier=False)
    if model is None:
        assert model_name != ''
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



code/scripts/lstm_alone.py [181:238]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                    ls = icsl_loss_function(intent_pred, slot_pred, intent_label, slot_label, valid_length - 2).mean()

                    if accumulate:
                        ls = ls / accumulate
                    ls.backward()
                    step_loss += ls.asscalar()

            # update
            if not accumulate or (batch_id + 1) % accumulate == 0:
                trainer.allreduce_grads()
                nlp.utils.clip_grad_global_norm(params, 1)
                trainer.update(1, ignore_stale_grad=True)

            if (batch_id + 1) % log_interval == 0:
                toc = time.time()
                # update metrics
                ic_metric.update([intent_label], [intent_pred])
                sl_metric.update(*process_seq_labels(slot_label, slot_pred, ignore_id=pad_label_id))
                log.info('Epoch: {}, Batch: {}/{}, speed: {:.2f} samples/s, lr={:.7f}, loss={:.4f}, intent acc={:.3f}, slot acc={:.3f}'
                         .format(epoch_id,
                                 batch_id,
                                 len(train_dataloader),
                                 log_num / (toc - tic),
                                 trainer.learning_rate,
                                 step_loss / log_interval,
                                 ic_metric.get()[1],
                                 sl_metric.get()[1]))

                tic = time.time()
                step_loss = 0
                log_num = 0

        mx.nd.waitall()
        epoch_toc = time.time()
        log.info('Time cost: {:.2f} s, Speed: {:.2f} samples/s'
                 .format(epoch_toc - epoch_tic, total_num/(epoch_toc - epoch_tic)))

        # evaluate on development set
        log.info('Evaluate on development set:')
        intent_acc, slot_f1 = evaluate(model=model, eval_input=dev_input)
        if slot_f1 > best_score[1]:
            best_score = (intent_acc, slot_f1)
            model.save_parameters(os.path.join(model_dir, model_name + '.params'))


def evaluate(model=None, model_name='', eval_input=''):
    """Evaluate the model on validation dataset.
    """
    ## Load model
    bert, vocabulary = nlp.model.get_model('bert_12_768_12',
                                           dataset_name='wiki_multilingual_uncased',
                                           pretrained=True,
                                           ctx=ctx,
                                           use_pooler=False,
                                           use_decoder=False,
                                           use_classifier=False)
    if model is None:
        assert model_name != ''
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -