code/scripts/bert_alone.py [298:348]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        intent_pred = intent_pred.asnumpy()
        slot_pred = slot_pred.asnumpy()
        valid_length = valid_length.asnumpy()

        for eid, y_intent, y_slot, length in zip(example_ids, intent_pred, slot_pred, valid_length):
            eid = eid.asscalar()
            length = int(length) - 2
            intent_id = y_intent.argmax(axis=-1)
            slot_ids = y_slot.argmax(axis=-1).tolist()[:length]
            slot_names = [idx2label[idx] for idx in slot_ids]
            merged_slot_names = merge_slots(slot_names, dev_alignment[eid] + [length])
            if eid not in all_results:
                all_results[eid] = _Result(intent_id, merged_slot_names)

    example_ids, utterances, labels, intents = load_tsv(eval_input)
    pred_intents = []
    label_intents = []
    for eid, intent in zip(example_ids, intents):
        label_intents.append(label2index(intent2idx, intent))
        pred_intents.append(all_results[eid].intent)
    intent_acc = sklearn.metrics.accuracy_score(label_intents, pred_intents)
    log.info("Intent Accuracy: %.4f" % intent_acc)

    pred_icsl = []
    label_icsl = []
    for eid, intent, slot_labels in zip(example_ids, intents, labels):
        label_icsl.append(str(label2index(intent2idx, intent)) + ' ' + ' '.join(slot_labels))
        pred_icsl.append(str(all_results[eid].intent) + ' ' + ' '.join(all_results[eid].slot_labels))
    exact_match = sklearn.metrics.accuracy_score(label_icsl, pred_icsl)
    log.info("Exact Match: %.4f" % exact_match)

    with open(conll_prediction_file, "w") as fw:
        for eid, utterance, labels in zip(example_ids, utterances, labels):
            preds = all_results[eid].slot_labels
            for w, l, p in zip(utterance, labels, preds):
                fw.write(' '.join([w, l, p]) + '\n')
            fw.write('\n')
    proc = subprocess.Popen(["perl", "conlleval.pl"], stdin=subprocess.PIPE, stdout=subprocess.PIPE)
    with open(conll_prediction_file) as f:
        stdout = proc.communicate(f.read().encode())[0]
    result = stdout.decode('utf-8').split('\n')[1]
    slot_f1 = float(result.split()[-1].strip())
    log.info("Slot Labeling: %s" % result)
    return intent_acc, slot_f1

# extract labels
train_input = data_dir + 'atis_train.tsv'
intent2idx, label2idx = get_label_indices(train_input)

for lang in ['EN', 'ES', 'DE', 'ZH', 'JA', 'PT', 'FR', 'HI', 'TR']:
    log.info('Train on %s:' % lang)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



code/scripts/lstm_alone.py [294:346]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        intent_pred = intent_pred.asnumpy()
        slot_pred = slot_pred.asnumpy()
        valid_length = valid_length.asnumpy()

        for eid, y_intent, y_slot, length in zip(example_ids, intent_pred, slot_pred, valid_length):
            eid = eid.asscalar()
            length = int(length) - 2
            intent_id = y_intent.argmax(axis=-1)
            slot_ids = y_slot.argmax(axis=-1).tolist()[:length]
            slot_names = [idx2label[idx] for idx in slot_ids]
            merged_slot_names = merge_slots(slot_names, dev_alignment[eid] + [length])
            if eid not in all_results:
                all_results[eid] = _Result(intent_id, merged_slot_names)

    example_ids, utterances, labels, intents = load_tsv(eval_input)
    pred_intents = []
    label_intents = []
    for eid, intent in zip(example_ids, intents):
        label_intents.append(label2index(intent2idx, intent))
        pred_intents.append(all_results[eid].intent)
    intent_acc = sklearn.metrics.accuracy_score(label_intents, pred_intents)
    log.info("Intent Accuracy: %.4f" % intent_acc)

    pred_icsl = []
    label_icsl = []
    for eid, intent, slot_labels in zip(example_ids, intents, labels):
        label_icsl.append(str(label2index(intent2idx, intent)) + ' ' + ' '.join(slot_labels))
        pred_icsl.append(str(all_results[eid].intent) + ' ' + ' '.join(all_results[eid].slot_labels))
    exact_match = sklearn.metrics.accuracy_score(label_icsl, pred_icsl)
    log.info("Exact Match: %.4f" % exact_match)

    with open(conll_prediction_file, "w") as fw:
        for eid, utterance, labels in zip(example_ids, utterances, labels):
            preds = all_results[eid].slot_labels
            for w, l, p in zip(utterance, labels, preds):
                fw.write(' '.join([w, l, p]) + '\n')
            fw.write('\n')
    proc = subprocess.Popen(["perl", "conlleval.pl"], stdin=subprocess.PIPE, stdout=subprocess.PIPE)
    with open(conll_prediction_file) as f:
        stdout = proc.communicate(f.read().encode())[0]
    result = stdout.decode('utf-8').split('\n')[1]
    slot_f1 = float(result.split()[-1].strip())
    log.info("Slot Labeling: %s" % result)
    return intent_acc, slot_f1


# extract labels
train_input = data_dir + 'atis_train.tsv'
intent2idx, label2idx = get_label_indices(train_input)

# Train
for lang in ['EN', 'ES', 'DE', 'ZH', 'JA', 'PT', 'FR', 'HI', 'TR']:
    log.info('Train on %s:' % lang)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -