JaheimLee

wenetspeech streaming ASR failed

3

When upgrade paddlespeech to the latest develop version, there are shape mismatch errors in logs. But the older version works well. ``` [2022-08-02 20:46:34,856] [ INFO] - we will use...

Can I get timestamp info by GPU inference?

5

I found timestamp info from [here](https://wenet.org.cn/wenet/python_binding.html?highlight=timestamp) . But it's only for cpu. Is it possible to get timestamp from GPU inference? For example by using your [docker server](https://github.com/wenet-e2e/wenet/tree/main/runtime/GPU)

关于自然语言理解任务的问题

2

Hi，我想和你们确认个问题。Huggingface的模型在文本分类任务上用BertForSequenceClassification这个类时，其中用到的是bert的pooled_output结果，然后接最终的一层classifier输出。而你们论文中说：“We build the downstream models for the natural language understanding tasks by adding a linear classifier on top of the “[CLS]" token to predict label probabilities.”。这个意思是仅用bert的CLS token，然后直接到最终的classifier是吗？因为我看你们预训练任务中有NSP任务，所以想确认一下文本分类你们具体用的哪种方式。谢谢~

streaming asr GPU利用率不到4%

1

本地用流式识别方式跑个长时间的音频文件，GPU利用率一直很低，怎么可以提升利用率啊？

通义模型

2

Hi，你们在clue上提交的通义模型对应的是PLUG吗？

can't install k2-sherpa

9

I installed all other dependencies using conda. And when I install sherpa using `pip install --verbose k2-sherpa`, I encountered the errors like https://github.com/k2-fsa/k2/issues/1031 . ``` [ 61%] Linking CXX shared...

RuntimeError: received 0 items of ancdata

1

When using prepare.sh file in wenetspeech from stage 1, a runtime error raised during computer features of `data/fbank/cuts_TEST_NET_raw.jsonl.gz`. Here is the logs: ``` 2022-08-02 22:39:27,066 INFO [compute_fbank_wenetspeech_dev_test.py:63] Loading data/fbank/cuts_TEST_NET_raw.jsonl.gz 2022-08-02...

position_ids with left padding

4

Hi, I have a question. According to the [huggingface code](https://github.com/huggingface/transformers/blob/151425ddb29d4ad1a121e8cce62000a2ac52d3ba/src/transformers/models/llama/modeling_llama.py#L522), `position_ids` is from 0 to seq_length. But if the padding side is left, text position doesn't start from 0. Will...

meta_instruction问题

1

如果说调用plugin插件需要额外的prompt信息是可以理解的。但是普通对话为什么要引入meta_instruction？即便想靠prompt引导模型输出正向结果，这个encode之后长度有200多是不是太浪费了？

RoPE bias

4

llama应该默认没有启用bias项。但按照苏神最新思路，把q,k的bias项加回来可以明显提升长度外推性能，作者考虑预训练测试一下不 https://kexue.fm/archives/9577