Differences

This shows you the differences between two versions of the page.

--- trans_learn:2021projects [2021/10/14 06:08] – [Meeting Notes] wuyr
+++ trans_learn:2021projects [2022/01/05 21:59] (current) – yang
@@ Line 1: / Line 1: @@
 ====== 2021 Project Page ======
+===== Transfer learning for battery parameter estimation =====
 ==== Meeting Notes ====
@@ Line 7: / Line 7: @@
   * Meeting summary:
         * Yanru: on reading: Unsupervised Domain Adaptation for Semantic Segmentation via CBST
-               TODO: make a plan of future work;
+        * Hanbing
-                     deeper understanding of papers/ better presentation of work
+          * research: optimal transport and domain adaptation problem
+          * reading: Joint distribution optimal transportation for domain adaptation; Few-Shot Cross Domain Battery Capacity Estimation
+          * slides: {{ :trans_learn:optimal_transport_and_domain_adaptation.pptx |optimal transport and domain adaptation}}
+  * TODO:
+        * Yanru: make a plan of future work; deeper understanding of papers/ better presentation of work
+        * Hanbing: do some research on the application of adversarial network to domain adaptation problem
@@ Line 14: / Line 20: @@
   * Attendee: Hanbing, Yanru, Yang
   * Meeting summary:
+        * Yanru:
+          * reading: transferable semi-supervised semantic segmentation; Adversarial Examples for Semantic Segmentation and Object Detection
+          * experiment: tried MIoU metric for H-score evaluation; on coding new dataset
+        * Hanbing:
+          * reading: Adversarial Discriminative Domain Adaptation
+          * research: Adversarial network in Domain Adaptation
+          * slides: {{ :trans_learn:adversarial_domain_adaptation.pptx |adversarial domain adaptation}}
+  * TODO:
+        * Yanru:
+          * more relative reading on transfer learning & U-net (also maybe some survey on semantic segmentation & related area);
+          * experiment redesign
+            * target and source setting (H-score more suitable for task transfer)
+            * dataset with fewer noise & lower difficulty
+            * new calculate method for H-score (devision depending on label instead of pixel location)
+        * Hanbing:
+          * deeper reading ADDA to find its advantage over other adversarial domain adaptation works
+          * modify the ADDA model to deal with battery features
+== 2021/10/20 ==
+  * Attendee: Hanbing, Yanru, Yang
+  * Meeting summary:
+        * Yanru:
+          * experiment: H-score trial on task transfer using ADE20K  -- especially good results for 'sky' task
+        * Hanbing:
+          * experiment: ADDA using battery pulse data
+          * slides:{{ :trans_learn:adda_test.pptx |}}
+  * TODO:
+        * Yanru:
+          * recode the experiment with pytorch
+          * cityscape dataset preprocessing
+          * change of target data size
+          * reading
+        * Hanbing:
+          * change convolution layer to fully connect or conv1D
+          * transform battery data into frequency domain
+== 2021/11/10 ==
+  * Attendee: Hanbing, Yang
+  * Meeting summary:
+    * Hanbing:
+      * experiment: unsupervised fully connection ADDA using battery pulse data
+  * TODO:
+    * Hanbing:
+      * experiment:
+        - source only: divide source dataset into training set and testing set to see if model is overfitting
+        - target only: using only target data to get target baseline model
+        - transfer learning: finetuning target feature extractor and source classifier useing a few target label
+== 2021/11/14 ==
+  * Attendee: Yanru, Yang
+  * Meeting summary:
+    - A brief review of several papers related to spatial-temporal data analyses ({{ :trans_learn:project_reading.pdf |}})
+    - Notes: {{:trans_learn:notes1114.png?linkonly|}}
+  * TODO: write background part of project plan
+== 2021/11/17 ==
+  * Attendee: Hanbing, Yang
+  * Meeting summary:
+    * Hanbing:
+      * experiment:
+        - source only and target only on original net and simplified net
+        - add penalty term on target model
+  * TODO:
+    * Hanbing:
+      * experiment:
+        - finetunning target model useing a few target label
+        - change visual form of result