superhg

Hi, @yl4579 here is my training loss curve and eval loss curve, convert result is not good. The loss curve indicates overfitting?  

the difference between any-to-many and any-to-any using multi-speakers is whether to use speaker-encoder?

training code for conformer + ctc missed files

1

hello,我对比了工程里的conformer实现代码，发现了一些与espnet对不上的，排除了一些版本的问题，发现有部分代码差别很大，如果直接用espnet最新的代码来训练conformer + ctc,有需要修改的地方吗？尤其是那个subsample这个地方

[BUG]:

4

### run gemini example failed `when run gemini example demo, below error msg occurs: [W socket.cpp:601] [c10d] The client socket has failed to connect to [::ffff:10.19.49.102]:35027 (errno: 110 - Connection...

finetune数据样本不需要增加[MASK]标记吗？

7

finetune Belle数据集的时候遇到了一个问题： ` File "/tal-vePFS/LLM/hegang/workspace/ChatGLM-chinese-insturct/modeling_chatglm.py", line 836, in forward mask_position = seq.index(mask_token) ValueError: 150000 is not in list 1%| | 1448/203736 [48:57

这里数据集里面 self.num_samples = 1000 * self.ds_len 为什么乘1000？ `class BlockDataset(data.Dataset): def __init__(self, ds, tokenizer, max_seq_len=1024, sample_across_doc=True, non_sentence_start=0.0, filter_english=False, **kwargs): """ sentence_start: the stripped article must start with a complete sentence """...

ValueError: 150000 is not in list

5

使用belle数据训练的时候，遇到这个错误，看了一下是训练文本文本中含有150000这个数字，出现了很多次。 ` File "/tal-vePFS/LLM/hegang/workspace/ChatGLM-chinese-insturct/modeling_chatglm.py", line 836, in forward mask_position = seq.index(mask_token) ValueError: 150000 is not in list `