“2024-01-06”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
第107行: 第107行:
 
|Tianhao Wang
 
|Tianhao Wang
 
||
 
||
*  
+
* Try using attention instead of FiLM:
 +
** validation loss: query emb cross-attn: -8.065 vs. query feature cross-attn: -9.127 vs. FiLM: -10.054
 +
** read a paper show that FiLM is better than attention in sound sep task
 +
* Huawei project testing
 
||
 
||
 
*
 
*

2025年1月6日 (一) 10:50的版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
Lantian Li
Ying Shi
Zhenghai You
Junming Yuan
Xiaolou Li
Zehua Liu
  • Organize my previous Experiment Result[1]
  • Add correct loss, but code encounters some bugs, so i fix it and train again(still training)
Pengqi Li
Wan Lin
  • NS: multi-enroll+margin bce: 1.638% -> 1.521% (5w spk)
Tianhao Wang
  • Try using attention instead of FiLM:
    • validation loss: query emb cross-attn: -8.065 vs. query feature cross-attn: -9.127 vs. FiLM: -10.054
    • read a paper show that FiLM is better than attention in sound sep task
  • Huawei project testing
Xiaoxue Luo
  • final exam
  • wrote the remaining course papers
Zhenyu Zhou
Junhui Chen
  • finished all course homework and final exams
Jiaying Wang
Yu Zhang
Wenqiang Du
Yang Wei
Turi
Yue Gu
  • report my recent work
  • read paper about functionally invariant path, which was reported by Prof. Wang.
Qi Qu
  • KWS:
    • pre-prod routine work for zh48+yichang5 models.
    • android demo backed with gRPC services deployed on peacock01.
  • QAT: exploration with little progress.