“2026-04-27”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（11位用户的14个中间修订版本未显示）
第32行： 第32行：
 
|Wenqiang Du
 
|Wenqiang Du
 
||
 
||
−
*
+
* Spech separation task
 +
** add some loss to reduce leakage problem
 +
* Baseline evaluation for audio event tasks
 
||
 
||
 
*
 
*
第43行： 第45行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*
+
* Train a audio separation model with larger dataset and remove speech subset with label problem
 
||
 
||
 
*
 
*
第76行： 第78行：
 
|Lily
 
|Lily
 
||
 
||
−
*
+
* GPT image generation: covers for 16 AI textbooks + AI100 (4 themes); collected feedback for further improvement
 +
* Reviewed Overleaf version of the 16 AI textbooks
 
||
 
||
 
*
 
*
第87行： 第90行：
 
|Pengqi Li
 
|Pengqi Li
 
||
 
||
−
*
+
* Continued the revision of the paper submitted to Journal of Chinese Information Processing
 
||
 
||
 
*
 
*
第110行： 第113行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*
+
* GPU Util: [https://z1et6d3xtb.feishu.cn/wiki/XX4NwX3tJiBDcgkMi0hcFUtInHh]
 +
* Paper:
 +
** Finish all lazy bubble prune accuracy & token consumption test
 +
** Refine paper with @chenjunhui
 +
* GAIA test (real word ToT MAS pipeline with function calling RAG)
 +
** original ACC: 11 / 52
 +
** ACC without all WebSearch: 8 / 52 (mostly are questions that can be solved by FileAnalyze or without any additional information)
 +
** ACC without WebSearch that seg-ECS < 0.4: 11 / 52
 +
** ACC without WebSearch that seg-ECS > 0.4: 8 / 52
 
||
 
||
 
*
 
*
第131行： 第142行：
  
 
|-
 
|-
−
|Jiaying Wang
+
|Xiaoxue Luo
 
||
 
||
−
*
+
* retrain the USS-eda model using the data and training configuration from the SJTU 's paper
 +
* train 2-3mix USS-tda model to compare the separation results with USS-eda(in training)
 
||
 
||
 
*
 
*
第144行： 第156行：
 
|Bochao Hu
 
|Bochao Hu
 
||
 
||
−
*
+
* confirm cvs3's plan with @xiaolou, complete all label generation, data part is done
 +
* handover gonganbu's tech report
 
||
 
||
 
*
 
*
第155行： 第168行：
 
|Hongcheng Zhang
 
|Hongcheng Zhang
 
||
 
||
−
*
+
* test Qwen models and write report
 +
* read some paper
 
||
 
||
 
*
 
*
第166行： 第180行：
 
|Weiman Sun
 
|Weiman Sun
 
||
 
||
−
*
+
* test Qwen models and write report
 +
* write graduation thesis
 
||
 
||
 
*
 
*
第174行： 第189行：
 
|Ge Gao
 
|Ge Gao
 
||
 
||
−
*
+
*experiments on audio separation and prepared ball data
 +
*wrote my graduation thesis
 
||
 
||
 
*
 
*
第183行： 第199行：
 
|Shuailong Li
 
|Shuailong Li
 
||
 
||
−
*
+
*Test on Teacher Wei's dataset
 +
*completed my graduation project UI and wrote the corresponding documentation.
 +
*Reproduction on the MTASS original dataset（loss：MSE+SI-SDR）
 +
**SDRi（speech=6.95 music=4.63 others=4.50 avg=5.43）
 +
**The perceived volume has not decreased
 +
 
 
||
 
||
 
*
 
*

2026年4月27日 (一) 11:00的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • Check response for 16 AI textbooks
  • Recheck micromagentics paper


Lantian Li
Wenqiang Du
  • Spech separation task
    • add some loss to reduce leakage problem
  • Baseline evaluation for audio event tasks
Yang Wei
  • Train a audio separation model with larger dataset and remove speech subset with label problem
Ying Shi
Yue Gu
  • write my phd thesis and revise one chapter
Lily
  • GPT image generation: covers for 16 AI textbooks + AI100 (4 themes); collected feedback for further improvement
  • Reviewed Overleaf version of the 16 AI textbooks
Pengqi Li
  • Continued the revision of the paper submitted to Journal of Chinese Information Processing
Junming Yuan
  • journal paper refinement(done and submit)
  • prepare ICASSP oral presentation(done)
Yu Zhang
  • GPU Util: [1]
  • Paper:
    • Finish all lazy bubble prune accuracy & token consumption test
    • Refine paper with @chenjunhui
  • GAIA test (real word ToT MAS pipeline with function calling RAG)
    • original ACC: 11 / 52
    • ACC without all WebSearch: 8 / 52 (mostly are questions that can be solved by FileAnalyze or without any additional information)
    • ACC without WebSearch that seg-ECS < 0.4: 11 / 52
    • ACC without WebSearch that seg-ECS > 0.4: 8 / 52
Junhui Chen
  • still sick
  • continue to refining the paper, and looking for references (80%)
Xiaoxue Luo
  • retrain the USS-eda model using the data and training configuration from the SJTU 's paper
  • train 2-3mix USS-tda model to compare the separation results with USS-eda(in training)
Bochao Hu
  • confirm cvs3's plan with @xiaolou, complete all label generation, data part is done
  • handover gonganbu's tech report
Hongcheng Zhang
  • test Qwen models and write report
  • read some paper
Weiman Sun
  • test Qwen models and write report
  • write graduation thesis
Ge Gao
  • experiments on audio separation and prepared ball data
  • wrote my graduation thesis
Shuailong Li
  • Test on Teacher Wei's dataset
  • completed my graduation project UI and wrote the corresponding documentation.
  • Reproduction on the MTASS original dataset（loss：MSE+SI-SDR）
    • SDRi（speech=6.95 music=4.63 others=4.50 avg=5.43）
    • The perceived volume has not decreased
取自“http://index.cslt.org/mediawiki/index.php?title=2026-04-27&oldid=44116”