Created by: jtyoui

请问一下: 我跑完这个代码: https://github.com/PaddlePaddle/ERNIE/blob/develop/demo/finetune_ner_dygraph.py
请问跑完之后怎么去使用呢??

我根据代码去使用得到的结果全是预测6, ['这', '也', '是', '美', '国', '、', '俄', '罗', '斯', '等', '石', '油', '、', '天', '然', '气', '都', '很', '丰', '富', '的', '国', '家', '搞', '代', '用', '燃', '料', '的', '原', '因'] [6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6] 很显然美国和俄罗斯都没有识别出现,

我不知道是不是我的测试代码有问题. 测试代码:

        with graph.guard():
            para_dict, _ = graph.load_dygraph(self.save_model)
            model = ErnieModelForTokenClassification.from_pretrained(
                self.pretreatment_model,
                num_labels=len(self.labels_index) + 1)
            model.set_dict(para_dict)
            place = fluid.CUDAPlace(0) if self.gpu else None
            with graph.base._switch_tracer_mode_guard_(is_train=False):
                model.eval()
                for step, (ids, sids, aligned_label, label, orig_pos) in enumerate(test_ds.start(place)):
                    chine, pre_labels = [], []
                    for id_ in ids.numpy():
                        chine.append(list(map(lambda x: self.index_vocab.get(x, 0), id_)))
                    loss, logic = model(ids, sids, labels=aligned_label)
                    for pos, lo, la in zip(orig_pos.numpy(), logic.numpy(), label.numpy()):
                        _dic = OrderedDict()
                        for p, l in zip(pos, lo):
                            _dic.setdefault(p, []).append(l)
                        del _dic[-100]
                        merged_lo = np.array([np.array(l).mean(0) for _, l in six.iteritems(_dic)])
                        merged_pred = np.argmax(merged_lo, -1)
                        la = la[np.where(la != -100)]
                        if len(la) > len(merged_pred):
                            print('accuracy loss due to truncation: label len:%d, truncate to %d' % (
                                len(la), len(merged_pred)))
                            merged_pred = np.pad(merged_pred, [0, len(la) - len(merged_pred)], mode='constant',
                                                 constant_values=-100)
                        pre_labels.append(merged_pred)
                    for c, l in zip(chine, pre_labels):
                        print(c)
                        print(l)
                        print()
                    return