Correlation Congruence for Knowledge Distillation
Motivation 目前大多數KD框架的pipeline 一張圖片送入teacher和student得到各自的feature和output 然後定義loss，拉進feature或者output的距離 但是他們都是直接在一個整圖級別上做的，student很難學習到teacher的那種處理類內和類間的能力 如下圖：白色空心圈是期望學習到的效果 左邊是期望類間可以拉開，右邊是期望類內可以聚攏 Meth
相關文章
- 1. Awesome Knowledge-Distillation
- 2. Knowledge Distillation
- 3. Structured Knowledge Distillation for Semantic Segmentation
- 4. Knowledge Distillation for Segmentation 筆記
- 5. knowledge distillation 論文閱讀之：Triplet Loss for Knowledge Distillation
- 6. knowledge distillation論文閱讀之：Learning from a Lightweight Teacher for Efficient Knowledge Distillation
- 7. Knowledge Distillation via Route Constrained Optimization
- 8. Knowledge Distillation 筆記
- 9. Tutorial: Knowledge Distillation
- 10. Similarity-Preserving Knowledge Distillation
- 更多相關文章...
- • Swift for 循環 - Swift 教程
- • Scala for循環 - Scala教程
- • PHP開發工具
- • C# 中 foreach 遍歷的用法
相關標籤/搜索
每日一句
-
每一个你不满意的现在，都有一个你没有努力的曾经。
最新文章
歡迎關注本站公眾號,獲取更多信息
相關文章
- 1. Awesome Knowledge-Distillation
- 2. Knowledge Distillation
- 3. Structured Knowledge Distillation for Semantic Segmentation
- 4. Knowledge Distillation for Segmentation 筆記
- 5. knowledge distillation 論文閱讀之：Triplet Loss for Knowledge Distillation
- 6. knowledge distillation論文閱讀之：Learning from a Lightweight Teacher for Efficient Knowledge Distillation
- 7. Knowledge Distillation via Route Constrained Optimization
- 8. Knowledge Distillation 筆記
- 9. Tutorial: Knowledge Distillation
- 10. Similarity-Preserving Knowledge Distillation