感谢公开源码！

我使用放出的plome模型，使用放出的训练集微调，使用放出的测试集（542条），最后的结果是 pinyin result: token check: p=0.933, r=0.874, f=0.902 token correction: p=0.975, r=0.852, f=0.909 fusion result: token check: p=0.936, r=0.871, f=0.902 token correction: p=0.966, r=0.841, f=0.899 sent check: p=0.867, r=0.800, f=0.832 sent correction: p=0.841, r=0.776, f=0.807 不知是否正确？

论文给出的结果为

不过测试集好像是1100条

这两个结果是同一种评价方式吗？

求解答

你好，请教一下关于预测部分代码实现的问题： ① 根据 train_eval_tagging.py，从中抽取了evaluate() 部分的代码进行修改。 ② 在实现predict过程中，由于测试样本没有标签，剔除了bert_tagging.py中关于label相关的代码，在调用create_model()方法时，出现了一些Error，请问在create_model()之前是否需要添加 graph=tf.get_default_graph() ? 还有为啥我的中间输出了一些类似这样的数值： (?, 40, 768) , (?, 4, 32) ，这里的 ？是不是表示输出有问题？

作者，您好，请问cbert是最终训练好的模型吗?评测直接使用cbert,会报错 Key loss/output_py_bias not found in checkpoint Not found: Key py_emb/GRU/rnn/gru_cell/candidate/bias not found in checkpoint

应该是用plome那个模型

您好，我在执行您的程序的时候遇到了 Traceback (most recent call last): File "G:/PLOME-main/finetune_src/train_eval_tagging.py", line 268, in print('\tsk_or_py=', FLAGS.sk_or_py) File "G:/PLOME-main/finetune_src/train_eval_tagging.py", line 203, in train start_time = time.time() File "G:\PLOME-main\venv\lib\site-packages\tensorflow\python\client\session.py", line 950, in run run_metadata_ptr) File "G:\PLOME-main\venv\lib\site-packages\tensorflow\python\client\session.py", line 1173, in _run feed_dict_tensor, options, run_metadata) File "G:\PLOME-main\venv\lib\site-packages\tensorflow\python\client\session.py", line 1350, in _do_run run_metadata) File "G:\PLOME-main\venv\lib\site-packages\tensorflow\python\client\session.py", line 1370, in _do_call raise type(e)(node_def, op, message) tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found. (0) Invalid argument: Key: stroke_ids. Can't parse serialized Example. [[{{node ParseSingleExample/ParseSingleExample}}]] [[IteratorGetNext]] [[IteratorGetNext/_475]] (1) Invalid argument: Key: stroke_ids. Can't parse serialized Example. [[{{node ParseSingleExample/ParseSingleExample}}]] [[IteratorGetNext]] 0 successful operations. 0 derived errors ignored. 请问这个问题您有解决方案么，本地环境和您在readme里面说的一样，谢谢您

def mask_process(self, input_sample):
    valid_ids = [idx for (idx, v) in enumerate(input_sample) if v not in self.invalid_ids]
    masked_sample = copy.deepcopy(list(input_sample))
    seq_len = len(masked_sample)
    masked_flgs = [0] * seq_len
    n_masked = int(len(valid_ids) * self.config['global_rate'])
    if n_masked < 1:
        n_masked = 1
    random.shuffle(valid_ids)
    for pos in valid_ids[:n_masked]:
        method = self.get_mask_method()
        if method == 'pinyin':
            new_c = self.same_py_confusion.get_confusion_item_by_ids(input_sample[pos])
            if new_c is not None:
                masked_sample[pos] = new_c
                masked_flgs[pos] = 1
        elif method == 'jinyin':
            new_c = self.simi_py_confusion.get_confusion_item_by_ids(input_sample[pos])
            if new_c is not None:
                masked_sample[pos] = new_c
                masked_flgs[pos] = 1
        elif method == 'stroke':
            new_c = self.sk_confusion.get_confusion_item_by_ids(input_sample[pos]) 
            if new_c is not None:
                masked_sample[pos] = new_c
                masked_flgs[pos] = 1
        elif method == 'random':
            new_c = self.all_token_ids[random.randint(0, self.n_all_token_ids)]
            if new_c is not None:
                masked_sample[pos] = new_c
                masked_flgs[pos] = 1
        elif method == 'keep': 
            masked_flgs[pos] = 1
    masked_py_ids = [self.tokenid_pyid.get(x, 1) for x in masked_sample]  
    masked_sk_ids = [self.tokenid_skid.get(x, 1) for x in masked_sample] 
    return np.asarray(masked_sample, dtype=np.int32), np.asarray(masked_flgs, dtype=np.int32), np.asarray(masked_py_ids, dtype=np.int32), np.asarray(masked_sk_ids, dtype=np.int32)

pre_train_src/mask.py中Mask这个类中的代码，看起来对每个样本，都只mask前面的字。