xiaomoguhzz's picture
Add files using upload-large-folder tool
c9aee57 verified

่งฃ่€ฆ่’ธ้ฆ vs ้›†ๆˆ่’ธ้ฆ ๆถˆ่žๅฎž้ชŒ่ฎพ่ฎก

ๆœ€ๅŽๆ›ดๆ–ฐ: 2026-01-11


๐Ÿ“‹ ๅฎก็จฟไบบ่ฆๆฑ‚

The advantages of the decoupled learning frameworkโ€”especially compared to integrated or joint distillation strategiesโ€”require clearer justification through more focused ablation studies.

DeCLIP claims decoupling avoids optimization conflicts, but lacks: (1) direct comparison with integrated distillation (e.g., joint content+context distillation without decoupling) in terms of mIoU/mAP and training stability


๐ŸŽฏ ๆ ธๅฟƒๅฏนๆฏ”ๅฎšไน‰

1. ่งฃ่€ฆ่’ธ้ฆ (Decoupled Distillation) - DeCLIP

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚                              DeCLIP ่งฃ่€ฆ่’ธ้ฆๆžถๆž„                                      โ”‚
โ”‚                                                                                     โ”‚
โ”‚  Input Image                                                                        โ”‚
โ”‚       โ†“                                                                             โ”‚
โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”               โ”‚
โ”‚  โ”‚                    EVA-CLIP / OpenAI-CLIP ViT                   โ”‚               โ”‚
โ”‚  โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  Block 0 โ†’ Block 1 โ†’ ... โ†’ Block (N-2)                  โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚            (ๅฎŒๆ•ด็š„ Attention + Residual + MLP)           โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜    โ”‚               โ”‚
โ”‚  โ”‚                             โ†“                                   โ”‚               โ”‚
โ”‚  โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚              Block (N-1) - ๆœ€ๅŽไธ€ไธช Block                โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚  LayerNorm โ†’ Q, K, V Projection                 โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚                                                 โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚      โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”        โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”               โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚      โ”‚   Q    โ”‚        โ”‚   V    โ”‚               โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚      โ””โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”˜        โ””โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”˜               โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚           โ”‚                 โ”‚                   โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚      Q @ Q^T           V @ V^T                  โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚           โ”‚                 โ”‚                   โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚      softmax            softmax                 โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚           โ†“                 โ†“                   โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚   โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”            โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚   โ”‚ Q Self-Sim   โ”‚  โ”‚ V Self-Sim   โ”‚            โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚   โ”‚  (Content)   โ”‚  โ”‚  (Context)   โ”‚            โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚   โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜            โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚          โ”‚                 โ”‚                    โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ”‚  โŒ ๆ— ๆฎ‹ๅทฎ่ฟžๆŽฅ   โŒ ๆ—  MLP                       โ”‚    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜    โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜    โ”‚               โ”‚
โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜               โ”‚
โ”‚                             โ†“                         โ†“                             โ”‚
โ”‚                    โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”      โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”                   โ”‚
โ”‚                    โ”‚  L_content Loss  โ”‚      โ”‚  L_context Loss  โ”‚                   โ”‚
โ”‚                    โ”‚  (Q features)    โ”‚      โ”‚  (V self-sim)    โ”‚                   โ”‚
โ”‚                    โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜      โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜                   โ”‚
โ”‚                             โ†“                         โ†“                             โ”‚
โ”‚                    Teacher: CLIP crops       Teacher: DINOv2 corr                   โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜

ๅ…ณ้”ฎ็‰น็‚น๏ผš

  • ๅœจๆœ€ๅŽไธ€ไธช Block ไฝฟ็”จ forward_without_rcffn ๆ–นๆณ•
  • ๅŽปๆމๆฎ‹ๅทฎ่ฟžๆŽฅ๏ผš้ฟๅ…็‰นๅพๆททๅˆ
  • ๅŽปๆމ MLP๏ผšไฟๆŒ Q ๅ’Œ V ็š„็บฏๅ‡€่กจ็คบ
  • Q ็š„่‡ช็›ธไผผๆ€ง โ†’ Content Loss๏ผˆไธŽ CLIP Teacher ๅฏน้ฝ๏ผ‰
  • V ็š„่‡ช็›ธไผผๆ€ง โ†’ Context Loss๏ผˆไธŽ DINOv2 ็›ธๅ…ณๆ€งๅฏน้ฝ๏ผ‰

2. ้›†ๆˆ่’ธ้ฆ (Integrated Distillation) - Baseline

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚                            Integrated ้›†ๆˆ่’ธ้ฆๆžถๆž„                                    โ”‚
โ”‚                                                                                     โ”‚
โ”‚  Input Image                                                                        โ”‚
โ”‚       โ†“                                                                             โ”‚
โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”               โ”‚
โ”‚  โ”‚                    EVA-CLIP / OpenAI-CLIP ViT                   โ”‚               โ”‚
โ”‚  โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚  Block 0 โ†’ Block 1 โ†’ ... โ†’ Block (N-1)                  โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚     ๅฎŒๆ•ด็š„ Attention + Residual + MLP๏ผˆๆ‰€ๆœ‰ๅฑ‚๏ผ‰          โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜    โ”‚               โ”‚
โ”‚  โ”‚                             โ†“                                   โ”‚               โ”‚
โ”‚  โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚                   Final Processing                       โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ”‚          Head (Linear) โ†’ LayerNorm                       โ”‚    โ”‚               โ”‚
โ”‚  โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜    โ”‚               โ”‚
โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜               โ”‚
โ”‚                             โ†“                                                       โ”‚
โ”‚                    โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”                                             โ”‚
โ”‚                    โ”‚  Final Features  โ”‚                                             โ”‚
โ”‚                    โ”‚   (่žๅˆ็š„็‰นๅพ)    โ”‚                                             โ”‚
โ”‚                    โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜                                             โ”‚
โ”‚                             โ”‚                                                       โ”‚
โ”‚           โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”                                     โ”‚
โ”‚           โ†“                 โ†“                 โ†“                                     โ”‚
โ”‚  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”  โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”                               โ”‚
โ”‚  โ”‚ L_content    โ”‚  โ”‚ L_context    โ”‚  โ”‚ L_region     โ”‚                               โ”‚
โ”‚  โ”‚ (ROIๆฑ ๅŒ–)    โ”‚  โ”‚ (self-sim)   โ”‚  โ”‚ (ROI sim)    โ”‚                               โ”‚
โ”‚  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜  โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜                               โ”‚
โ”‚           โ†“                 โ†“                 โ†“                                     โ”‚
โ”‚  Teacher: CLIP       Teacher: DINOv2    Teacher: DINOv2                             โ”‚
โ”‚                                                                                     โ”‚
โ”‚              L_total = L_content + L_context + L_region                             โ”‚
โ”‚              (ๆ‰€ๆœ‰ Loss ไฝœ็”จไบŽๅŒไธ€ไธช่žๅˆ็‰นๅพ๏ผ)                                       โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜

ๅ…ณ้”ฎ็‰น็‚น๏ผš

  • ไฝฟ็”จๅฎŒๆ•ด็š„ ViT forward๏ผˆๅŒ…ๅซๆ‰€ๆœ‰ๆฎ‹ๅทฎ่ฟžๆŽฅๅ’Œ MLP๏ผ‰
  • ๆœ€็ปˆ่พ“ๅ‡บๆ˜ฏ ่žๅˆๅŽ็š„็‰นๅพ
  • Content ๅ’Œ Context Loss ้ƒฝไฝœ็”จไบŽ ๅŒไธ€ไธช่žๅˆ็‰นๅพ
  • ๅฏ่ƒฝๅฏผ่‡ด ๆขฏๅบฆๅ†ฒ็ช๏ผšไธคไธชไปปๅŠก็š„ๆขฏๅบฆๆ–นๅ‘ๅฏ่ƒฝ็›ธๅ

๐Ÿ“Š ๅฎž้ชŒๆ–นๆกˆ

ๅฎž้ชŒ 1๏ผš่ฎญ็ปƒๅŠจๆ€ๆ›ฒ็บฟ๏ผˆTraining Dynamics๏ผ‰

็›ฎ็š„๏ผš่ฏๆ˜Ž่งฃ่€ฆ่’ธ้ฆ่ฎญ็ปƒๆ›ด็จณๅฎš

่ฎฐๅฝ•ๆŒ‡ๆ ‡๏ผˆๆฏไธช iteration๏ผ‰๏ผš

  • L_content vs Iteration
  • L_context vs Iteration
  • L_total vs Iteration

้ข„ๆœŸ็ป“ๆžœ๏ผš

  • DeCLIP๏ผšๆ›ฒ็บฟๅนณ็จณไธ‹้™๏ผŒๆ”ถๆ•›ๅฟซ
  • Integrated๏ผšๆ›ฒ็บฟ้œ‡่ก๏ผŒๆ”ถๆ•›ๆ…ข๏ผŒๅฏ่ƒฝๅ‡บ็Žฐ Loss ๅๅผน

ๅฎž้ชŒ 2๏ผšๆขฏๅบฆๅ†ฒ็ชๅˆ†ๆž๏ผˆGradient Conflict Analysis๏ผ‰

็›ฎ็š„๏ผšๆ•ฐๅญฆ่ฏๆ˜ŽไธคไธชไปปๅŠกๅญ˜ๅœจไผ˜ๅŒ–ๅ†ฒ็ช

ๆ–นๆณ•๏ผˆๅ‚่€ƒ PCGrad ่ฎบๆ–‡๏ผ‰๏ผš

# ๅœจ Integrated baseline ่ฎญ็ปƒๆ—ถ๏ผš
# 1. ๅˆ†ๅˆซ่ฎก็ฎ— L_content ๅ’Œ L_context ๅฏนๆœ€ๅŽ N ๅฑ‚็š„ๆขฏๅบฆ
grad_content = torch.autograd.grad(L_content, params, retain_graph=True)
grad_context = torch.autograd.grad(L_context, params, retain_graph=True)

# 2. ่ฎก็ฎ—ๆขฏๅบฆไฝ™ๅผฆ็›ธไผผๅบฆ
cos_sim = cosine_similarity(grad_content, grad_context)

# 3. ้€ๅฑ‚็ปŸ่ฎก
# cos_sim < 0: ๆขฏๅบฆๅ†ฒ็ช
# cos_sim โ‰ˆ 0: ๆขฏๅบฆๆญฃไบค
# cos_sim > 0: ๆขฏๅบฆไธ€่‡ด

ๅˆ†ๆž่Œƒๅ›ด๏ผš้€ๅฑ‚ๅˆ†ๆžๆฏไธช ViT Block

้ข„ๆœŸ็ป“ๆžœ๏ผš

  • ๆขฏๅบฆไฝ™ๅผฆ็›ธไผผๅบฆไธป่ฆๅˆ†ๅธƒๅœจ ่ดŸๅ€ผๅŒบ้—ด๏ผˆๅ†ฒ็ช๏ผ‰ๆˆ– ๆŽฅ่ฟ‘ 0๏ผˆๆญฃไบค๏ผ‰
  • ่ฏดๆ˜ŽไธคไธชไปปๅŠก็š„ๆขฏๅบฆๆ–นๅ‘็ปๅธธ็›ธๅๆˆ–ไธ็›ธๅ…ณ

ๅฎž้ชŒ 3๏ผšPCA ๅฏ่ง†ๅŒ–

็›ฎ็š„๏ผš็›ด่ง‚ๅฑ•็คบ่งฃ่€ฆๆ•ˆๆžœ

ๅฏนๆฏ”ๅ†…ๅฎน๏ผš

  • DeCLIP ่งฃ่€ฆ่’ธ้ฆ็š„ๆœ€็ปˆ่พ“ๅ‡บ็‰นๅพ
  • Integrated ้›†ๆˆ่’ธ้ฆ็š„ๆœ€็ปˆ่พ“ๅ‡บ็‰นๅพ

ๆ–นๆณ•๏ผš

  • ้€‰ๅ–ๆŒ‘ๆˆ˜ๆ€งๅ›พ็‰‡๏ผˆๅคš็‰ฉไฝ“ใ€ๅคๆ‚่ƒŒๆ™ฏ๏ผ‰
  • ๆๅ–็‰นๅพ็š„ๅ‰ 3 ไธปๆˆๅˆ†๏ผŒๆ˜ ๅฐ„ไธบ RGB

้ข„ๆœŸ็ป“ๆžœ๏ผš

  • Integrated๏ผš็‰นๅพ"ๆจก็ณŠ"ๆˆ–"ๆททไนฑ"
  • DeCLIP๏ผš่พน็•Œๆธ…ๆ™ฐ๏ผŒ่ฏญไน‰ๅ‡†็กฎ

ๅฎž้ชŒ 4๏ผšCKA ๅˆ†ๆž

็›ฎ็š„๏ผšๅˆ†ๆž็‰นๅพ็›ธไผผๆ€ง

ๅฏนๆฏ”ๅ†…ๅฎน๏ผš

  1. DeCLIP ็š„ Q ็‰นๅพ vs V ็‰นๅพ
  2. Integrated ่พ“ๅ‡บ็‰นๅพ vs DeCLIP Q ็‰นๅพ
  3. Integrated ่พ“ๅ‡บ็‰นๅพ vs DeCLIP V ็‰นๅพ

้ข„ๆœŸ็ป“ๆžœ๏ผš

  • DeCLIP ็š„ Q ๅ’Œ V ๆœ‰ไธ€ๅฎšๅทฎๅผ‚๏ผˆ่ฏๆ˜Ž็กฎๅฎž่งฃ่€ฆไบ†๏ผ‰
  • Integrated ่พ“ๅ‡บไธŽ DeCLIP Q/V ็š„ๅ…ณ็ณปๆจก็ณŠ

ๅฎž้ชŒ 5๏ผšๅฎš้‡่กจๆ ผๅฏนๆฏ”๏ผˆๆš‚ไธๅฎž็Žฐ๏ผŒ็”จๆˆท่‡ช่กŒ่ฏ„ไผฐ๏ผ‰

ๆŒ‡ๆ ‡๏ผš

  • Detection: mAP
  • Segmentation: mIoU

๐Ÿ”ง ่ฎญ็ปƒ้…็ฝฎ

็ปŸไธ€้…็ฝฎ

้…็ฝฎ้กน ๅ€ผ
VFM Teacher DINOv2
Epochs 6
Learning Rate 1e-5
Batch Size 16 (total)
Optimizer AdamW
Scheduler Cosine
Dataset COCO

ๆจกๅž‹้…็ฝฎ

Student Model VFM Size
EVA02-CLIP-B-16 DINOv2-B
EVA02-CLIP-L-14-336 DINOv2-L
OpenAI-CLIP-B-16 DINOv2-B
OpenAI-CLIP-L-14-336 DINOv2-L

Loss ้…็ฝฎ

DeCLIP๏ผˆ่งฃ่€ฆ๏ผ‰๏ผš

  • Content Loss ๆƒ้‡: 1.0
  • Context Loss ๆƒ้‡: 0.25
  • Region Loss ๆƒ้‡: 0.05

Integrated๏ผˆ้›†ๆˆ๏ผ‰๏ผš

  • Content Loss ๆƒ้‡: 1.0
  • Context Loss ๆƒ้‡: 1.0๏ผˆ็›ดๆŽฅ็›ธๅŠ ๏ผŒไธ็”จ็ณปๆ•ฐ๏ผ‰

๐Ÿ“ ไปฃ็ ็ป„็ป‡

decoupling_analysis/
โ”œโ”€โ”€ DESIGN.md                           # ๆœฌ่ฎพ่ฎกๆ–‡ๆกฃ
โ”œโ”€โ”€ PROGRESS.md                         # ่ฟ›ๅบฆ่ฎฐๅฝ•
โ”œโ”€โ”€ gradient_analysis.py                # ๆขฏๅบฆๅ†ฒ็ชๅˆ†ๆž
โ”œโ”€โ”€ pca_visualization.py                # PCA ๅฏ่ง†ๅŒ–
โ””โ”€โ”€ cka_analysis.py                     # CKA ๅˆ†ๆž

scripts/decoupling_ablation/
โ”œโ”€โ”€ dist_integrated_eva_vitb16_coco.sh      # EVA-B/16 ๅคšๅก่ฎญ็ปƒ
โ”œโ”€โ”€ dist_integrated_eva_vitL14_336_coco.sh  # EVA-L/14-336 ๅคšๅก่ฎญ็ปƒ
โ”œโ”€โ”€ dist_integrated_openai_vitb16_coco.sh   # OpenAI-B/16 ๅคšๅก่ฎญ็ปƒ
โ”œโ”€โ”€ dist_integrated_openai_vitL14_coco.sh   # OpenAI-L/14 ๅคšๅก่ฎญ็ปƒ
โ”œโ”€โ”€ debug_integrated_eva_vitb16_coco.sh     # EVA-B/16 ๅ•ๅก่ฐƒ่ฏ•
โ”œโ”€โ”€ debug_integrated_eva_vitL14_336_coco.sh # EVA-L/14-336 ๅ•ๅก่ฐƒ่ฏ•
โ”œโ”€โ”€ debug_integrated_openai_vitb16_coco.sh  # OpenAI-B/16 ๅ•ๅก่ฐƒ่ฏ•
โ””โ”€โ”€ debug_integrated_openai_vitL14_coco.sh  # OpenAI-L/14 ๅ•ๅก่ฐƒ่ฏ•

src/training/
โ”œโ”€โ”€ integrated_distill.py     # ้›†ๆˆ่’ธ้ฆๅฎž็Žฐ (IntegratedDistillation ็ฑป)
โ”œโ”€โ”€ main.py                   # (ๅทฒไฟฎๆ”น) ๆทปๅŠ  integrated ็‰ˆๆœฌๆ”ฏๆŒ
โ””โ”€โ”€ params.py                 # (ๅทฒไฟฎๆ”น) ๆทปๅŠ  --version integrated ้€‰้กน

๐Ÿ“ ๅฎž็Žฐๆญฅ้ชค

Step 1: ๅฎž็Žฐ Integrated Distillation โœ…

  • ๅˆ›ๅปบ src/training/integrated_distill.py
  • ไฝฟ็”จๅฎŒๆ•ด็š„ ViT forward๏ผˆmode="vanilla"๏ผŒไธ่ฐƒ็”จ forward_without_rcffn๏ผ‰
  • ๅœจๆœ€็ปˆ่พ“ๅ‡บ็‰นๅพไธŠๅŒๆ—ถ่ฎก็ฎ— Content + Context Loss

Step 2: ๅฎž็Žฐ Gradient Analysis โœ…

  • ๅˆ›ๅปบ gradient_analysis.py
  • ๅœจ่ฎญ็ปƒ่ฟ‡็จ‹ไธญ่ฎฐๅฝ•ๆขฏๅบฆไฝ™ๅผฆ็›ธไผผๅบฆ
  • ้€ๅฑ‚ๅˆ†ๆž

Step 3: ๅฎž็Žฐ PCA Visualization โœ…

  • ๅˆ›ๅปบ pca_visualization.py
  • ๅฏนๆฏ” DeCLIP vs Integrated ็š„็‰นๅพ

Step 4: ๅฎž็Žฐ CKA Analysis โœ…

  • ๅˆ›ๅปบ cka_analysis.py
  • ๅˆ†ๆž Q, V, ่พ“ๅ‡บ็‰นๅพ็š„็›ธไผผๆ€ง

Step 5: ๅˆ›ๅปบ่ฎญ็ปƒ่„šๆœฌ โœ…

  • EVA-CLIP B/L
  • OpenAI CLIP B/L

Step 6: ไปฃ็ ๅฎก้˜…ไธŽไฟฎๅค โœ… (2026-01-11)

  • ไฟฎๅค extract_roi_features_integrated ็ผบๅฐ‘ .transpose(-2,-1) ็š„็ปดๅบฆ้—ฎ้ข˜
  • ไฟฎๅค region_scd_loss ไธญไธๅฟ…่ฆ็š„ transpose
  • ็งป้™คๆœชไฝฟ็”จ็š„ๅฏผๅ…ฅ (autocast, math)

โš ๏ธ ๆณจๆ„ไบ‹้กน

  1. PCGrad ไป…็”จไบŽๅˆ†ๆž๏ผšไธไฝฟ็”จๅฎƒๆฅ่งฃๅ†ณๅ†ฒ็ช๏ผŒๅช็”จๆฅ่ฏๆ˜Žๅ†ฒ็ชๅญ˜ๅœจ
  2. Loss ็›ดๆŽฅ็›ธๅŠ ๏ผšIntegrated ็š„ Content ๅ’Œ Context Loss ็›ดๆŽฅ็›ธๅŠ ๏ผŒไธ่ฐƒๆƒ้‡
  3. ่ฎญ็ปƒ้…็ฝฎไธ€่‡ด๏ผš้™คไบ†่งฃ่€ฆ vs ้›†ๆˆ็š„ๅทฎๅผ‚๏ผŒๅ…ถไป–้…็ฝฎๅฎŒๅ…จไธ€่‡ด
  4. ่ฏ„ไผฐ็”จๆˆท่‡ช่กŒ๏ผš่ฎญ็ปƒๅฎŒๆฏ•ๅŽ็”จๆˆท่‡ชๅทฑ่ฟ›่กŒ mAP/mIoU ่ฏ„ไผฐ