“2024-07-29”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
 
(9位用户的11个中间修订版本未显示)
第6行: 第6行:
 
|Dong Wang
 
|Dong Wang
 
||
 
||
*
+
* AIGraph slides done
 +
* Check for thermal face recognition paper
 +
* Quick check for Guyue's paper
 +
 
 
||
 
||
 
*
 
*
第32行: 第35行:
 
|Ying Shi
 
|Ying Shi
 
||
 
||
*  
+
* Finish training U-Net based text-enroll keyword spotting
 +
* continue work on cohort conditional-chain [https://z1et6d3xtb.feishu.cn/docx/X6eLdaooUoVsCzxneGecUubInCc?from=from_copylink group work]
 
||
 
||
 
*
 
*
第44行: 第48行:
 
||
 
||
 
* Complete the Reproduce of IRA[https://z1et6d3xtb.feishu.cn/docx/TdlDdycVnoYNn3xI7QacnYrTnuc]
 
* Complete the Reproduce of IRA[https://z1et6d3xtb.feishu.cn/docx/TdlDdycVnoYNn3xI7QacnYrTnuc]
* Design a new TSE structure using U-NET and G&T Transformer (idea form sepreformer)
+
* Design a new TSE structure using U-NET and G&L Transformer (idea form sepreformer)
 
* Write the work content for Huawei's first phase as ICCIP2024
 
* Write the work content for Huawei's first phase as ICCIP2024
 
||
 
||
第59行: 第63行:
 
** the base model for the 1st iteration is finished on hawk02.
 
** the base model for the 1st iteration is finished on hawk02.
 
** the base model for the 2nd iteration need to migrate to dragon03(in progress)
 
** the base model for the 2nd iteration need to migrate to dragon03(in progress)
 +
** Beginner's Guide for pretraining Hubert with fairseq:[https://z1et6d3xtb.feishu.cn/docx/TGBPdRS8HoTHWdxGn7scbTqPnxd]
 
||
 
||
 
*
 
*
第80行: 第85行:
 
|Xiaolou Li
 
|Xiaolou Li
 
||
 
||
*
+
* LRS-30h PALR2 (4 epoch result)
 +
** VSR: 29.74%
 +
** Refinement: 29.45%
 +
* Calibration Test [https://z1et6d3xtb.feishu.cn/docx/CpnKdz2ruoVBxOx59wLcT9FYnSg?from=from_copylink]
 +
 
 
||
 
||
 
*
 
*
第104行: 第113行:
 
|Pengqi Li
 
|Pengqi Li
 
||
 
||
*
+
* Trianed a new attention pooling with condition(Analysis ongoing).[https://z1et6d3xtb.feishu.cn/docx/PgYpdmtH2oE1YexbDB8c5jW0nTh]
 
||
 
||
 
*
 
*
第128行: 第137行:
 
|Tianhao Wang
 
|Tianhao Wang
 
||
 
||
*
+
* SoundFilter weekly report
 
||
 
||
*
+
* SoundFilter reproduce
 
||
 
||
 
*
 
*
第139行: 第148行:
 
|Zhenyu Zhou
 
|Zhenyu Zhou
 
||
 
||
*
+
*Clip Norm results[https://z1et6d3xtb.feishu.cn/docx/AVKSdXxJooSgfXxbYW1c1lNrnab]
 
||
 
||
 
*
 
*
第163行: 第172行:
 
|Jiaying Wang
 
|Jiaying Wang
 
||
 
||
*
+
* paper reading(report today)
 +
* live broadcast
 
||
 
||
 
*
 
*
第196行: 第206行:
 
|Yang Wei
 
|Yang Wei
 
||
 
||
*
+
* AIBabel KWS
 +
** Prepare negative test data (from cn-celeb, aishell-4)
 
||
 
||
 
*
 
*

2024年7月29日 (一) 10:55的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • AIGraph slides done
  • Check for thermal face recognition paper
  • Quick check for Guyue's paper
Lantian Li
  • GPU status [1]
  • AI graph
    • Slides checking (50/50)
    • High school handbook (12/40)
  • High school handbook (20/40)
Ying Shi
  • Finish training U-Net based text-enroll keyword spotting
  • continue work on cohort conditional-chain group work
Zhenghai You
  • Complete the Reproduce of IRA[2]
  • Design a new TSE structure using U-NET and G&L Transformer (idea form sepreformer)
  • Write the work content for Huawei's first phase as ICCIP2024
Junming Yuan
  • Reimplementation of the Hubert baseline:
    • fix some bugs
    • the base model for the 1st iteration is finished on hawk02.
    • the base model for the 2nd iteration need to migrate to dragon03(in progress)
    • Beginner's Guide for pretraining Hubert with fairseq:[3]
Chen Chen
Xiaolou Li
  • LRS-30h PALR2 (4 epoch result)
    • VSR: 29.74%
    • Refinement: 29.45%
  • Calibration Test [4]
Zehua Liu
  • LRS3-30h: VSP-LLM - cluster(WER : 28.11%) < VSP-LLM (WER : 29.1%)
  • LRS3-30h: VSP-LLM - cluster + adaptive_mask(WER : 27.75%) < VSP-LLM (WER : 29.1%)
  • LRS3-30h: In-Context-learning (still training)
Pengqi Li
  • Trianed a new attention pooling with condition(Analysis ongoing).[5]
Wan Lin
  • Neural Scoring
    • First draft of paper finished [6]
    • Supplement experimental results
Tianhao Wang
  • SoundFilter weekly report
  • SoundFilter reproduce
Zhenyu Zhou
  • Clip Norm results[7]
Junhui Chen
  • Neural Scoring:
    • Paper Writing (1st Ver. finished with LW)
    • Supplement the experiments
Jiaying Wang
  • paper reading(report today)
  • live broadcast
Yu Zhang
Wenqiang Du
Yang Wei
  • AIBabel KWS
    • Prepare negative test data (from cn-celeb, aishell-4)
Lily
  • Prepare for high shcool summer trip class(last Sunday)
  • Accident & get sick
Turi
Yue Gu
  • complete and revise the DPR paper
Qi Qu
  • AED:
    • AudioSet data prepared.
    • Positive samples of "cries" collected and to be annotated.
  • KWS:
    • B6-based service optimized with memory consumption considerably reduced (~600MB v.s. formerly ~2GB).