27Stage1 runs
10Stage2 runs
23OT diagnostic runs
37total parsed logs
路线状态
- Phase 1: Stage1 cross-attention can be patched by OT-gated routing without changing the official checkpoint format.
- Phase 2: 2D masks are now soft OT priors and edge maps enter the cost, so segmentation errors no longer monopolize part assignment.
- Phase 3: A virtual patch channel gives occluded or weakly visible parts an explicit route to learned prior information.
- Phase 4: Stage2 now exposes spatial mask moments and a zero-initialized articulation residual adapter for joint prediction.
可视化







消融表
| Run | Phase | Steps | Loss | Art L2 | Gate | Edge delta | Virtual mass | Mask parts | Empty | Note |
|---|---|---|---|---|---|---|---|---|---|---|
Stage1 baseline smokepact_finetune_route_32_20260516/stage1_smoke | baseline | 1 | 0.03043 | - | - | - | - | - | - | Official-equivalent trainer, no OT router. |
OT first blockpact_stage1_ot_router_first_diag_smoke_20260520 | phase1 | 1 | 0.0304 | - | 0.981 | - | - | - | - | One Stage1 cross-attention block patched with OT gating. |
OT mask priorpact_stage1_ot_router_maskprior_smoke_20260520 | phase2 | 1 | 0.02921 | - | 0.2925 | - | - | - | - | Mask demoted from hard condition to OT cost prior. |
OT mask prior 5-steppact_stage1_ot_router_maskprior_5step_sgd1e8_20260520 | phase2 | 5 | 0.04972 | - | 0.2306 | - | - | - | - | SGD stability check for soft mask prior. |
OT RGB edgepact_stage1_ot_router_rgb_edge_1step_20260520 | phase2 | 1 | 0.02921 | - | 0.2925 | 0.001399 | - | - | - | RGB Sobel edge regularized cost. |
OT edge 5-steppact_stage1_ot_router_edge_fixed_5step_20260520 | phase2 | 5 | 0.04973 | - | 0.231 | 0.001319 | - | - | - | Mask-edge regularized 5-step stability check. |
OT virtual patchpact_stage1_ot_router_virtual005_rgb_fixedgate_5step_20260520 | phase3 | 5 | 0.03779 | - | 0.2607 | 0.00119 | 0.04999 | - | - | Virtual image token with visible/virtual gate normalization. |
OT first-thirdpact_stage1_ot_router_firstthird_1step_20260520 | phase1 | 1 | 0.02923 | - | 0.2901 | - | - | - | - | Patch first third of Stage1 cross-attention blocks. |
OT edge first-thirdpact_stage1_ot_router_edge_firstthird_fixed_1step_20260520 | phase2 | 1 | 0.02924 | - | 0.2905 | 0.001509 | - | - | - | First-third patch plus edge-regularized cost. |
Stage2 baseline smokepact_finetune_route_32_20260516/stage2_smoke | baseline | 1 | 3.2046 | 0.09492 | - | - | - | - | - | Official-equivalent Stage2 smoke. |
Stage2 spatial diagpact_stage2_spatial_diag_smoke_20260521 | phase4 | 1 | 2.3856 | 0.1107 | - | - | - | 5.0 | 1.0 | Mask-derived spatial moment diagnostics for articulation rows. |
Stage2 spatial adapter 1-steppact_stage2_spatial_adapter_smoke_20260521 | phase4 | 1 | 7.4112 | 0.1336 | - | - | - | 3.0 | 1.0 | Zero-initialized articulation residual adapter smoke. |
Stage2 spatial adapter 5-steppact_stage2_spatial_adapter_5step_20260521 | phase4 | 5 | 3.3670 | 0.1141 | - | - | - | 6.0 | 2.0 | Spatial adapter 5-step SGD stability check. |
Matrix Stage1 baselinepact_ot_matrix_20260521_stage1_baseline | matrix | 8 | 1.0631 | - | - | - | - | - | - | Quick matrix baseline with the official-equivalent Stage1 path. |
Matrix Stage1 OT firstpact_ot_matrix_20260521_stage1_ot_first | matrix | 8 | 1.0631 | - | 0.9751 | 0 | 0 | - | - | Quick matrix with one OT-gated cross-attention block. |
Matrix Stage1 mask priorpact_ot_matrix_20260521_stage1_mask_prior | matrix | 8 | 1.0631 | - | 0.3237 | 0 | 0 | - | - | Quick matrix with mask labels as a soft OT prior. |
Matrix Stage1 RGB edgepact_ot_matrix_20260521_stage1_rgb_edge | matrix | 8 | 1.0631 | - | 0.3237 | 0.001583 | 0 | - | - | Quick matrix with RGB Sobel edge-regularized OT cost. |
Matrix Stage1 virtual patchpact_ot_matrix_20260521_stage1_virtual_patch | matrix | 8 | 1.0334 | - | 0.1462 | 0.001448 | 0.04999 | - | - | Quick matrix with a virtual patch channel. |
Matrix Stage1 first-third edgepact_ot_matrix_20260521_stage1_firstthird_edge | matrix | 3 | 1.0303 | - | 0.3603 | 0.001898 | 0 | - | - | Quick matrix with OT injected into the first third of Stage1 blocks. |
Matrix Stage2 spatial diagpact_ot_matrix_20260521_stage2_spatial_diag | matrix | 5 | 3.4941 | 0.2005 | - | - | - | 4.0 | 2.0 | Quick matrix Stage2 spatial diagnostics. |
Matrix Stage2 spatial adapterpact_ot_matrix_20260521_stage2_spatial_adapter | matrix | 5 | 3.3674 | 0.1141 | - | - | - | 6.0 | 2.0 | Quick matrix Stage2 spatial adapter. |
Full Stage1 baselinepact_ot_matrix_20260521_full_stage1_baseline | full | 60 | 1.0311 | - | - | - | - | - | - | Full matrix baseline with final checkpoint export. |
Full Stage1 OT firstpact_ot_matrix_20260521_full_stage1_ot_first | full | 60 | 1.0311 | - | 0.9707 | 0 | 0 | - | - | Full matrix with one OT-gated cross-attention block. |
Full Stage1 mask priorpact_ot_matrix_20260521_full_stage1_mask_prior | full | 60 | 1.0311 | - | 0.271 | 0 | 0 | - | - | Full matrix with mask labels as a soft OT prior. |
Full Stage1 RGB edgepact_ot_matrix_20260521_full_stage1_rgb_edge | full | 60 | 1.0311 | - | 0.2712 | 0.001356 | 0 | - | - | Full matrix with RGB Sobel edge-regularized OT cost. |
Full Stage1 virtual patchpact_ot_matrix_20260521_full_stage1_virtual_patch | full | 60 | 1.0240 | - | 0.29 | 0.001643 | 0.04999 | - | - | Full matrix with a virtual patch channel. |
Full Stage1 first-third edgepact_ot_matrix_20260521_full_stage1_firstthird_edge | full | 20 | 1.0325 | - | 0.1809 | 0.002062 | 0 | - | - | Full matrix with OT injected into the first third of Stage1 blocks. |
Full Stage2 spatial diagpact_ot_matrix_20260521_full_stage2_spatial_diag | full | 40 | 1.5619 | 0.06205 | - | - | - | 10.0 | 3.0 | Full matrix Stage2 spatial diagnostics. |
Full Stage2 spatial adapterpact_ot_matrix_20260521_full_stage2_spatial_adapter | full | 40 | 5.9822 | 0.157 | - | - | - | 6.0 | 1.0 | Full matrix Stage2 spatial adapter. |
Official full Stage1 baselinepact_ot_matrix_official_20260521_full_stage1_baseline | official-full | 60 | 0.04795 | - | - | - | - | - | - | Official-initialized full matrix baseline with final checkpoint export. |
Official full Stage1 OT firstpact_ot_matrix_official_20260521_full_stage1_ot_first | official-full | 60 | 0.04791 | - | 0.9351 | 0 | 0 | - | - | Official-initialized full matrix with one OT-gated cross-attention block. |
Official full Stage1 mask priorpact_ot_matrix_official_20260521_full_stage1_mask_prior | official-full | 60 | 0.04789 | - | 0.2683 | 0 | 0 | - | - | Official-initialized full matrix with mask labels as a soft OT prior. |
Official full Stage1 RGB edgepact_ot_matrix_official_20260521_full_stage1_rgb_edge | official-full | 60 | 0.0479 | - | 0.2686 | 0.001104 | 0 | - | - | Official-initialized full matrix with RGB Sobel edge-regularized OT cost. |
Official full Stage1 virtual patchpact_ot_matrix_official_20260521_full_stage1_virtual_patch | official-full | 60 | 0.05485 | - | 0.2864 | 0.0009899 | 0.04999 | - | - | Official-initialized full matrix with a virtual patch channel. |
Official full Stage1 first-third edgepact_ot_matrix_official_20260521_full_stage1_firstthird_edge | official-full | 20 | 0.04858 | - | 0.1845 | 0.001843 | 0 | - | - | Official-initialized full matrix with OT injected into the first third of Stage1 blocks. |
Official full Stage2 spatial diagpact_ot_matrix_official_20260521_full_stage2_spatial_diag | official-full | 40 | 1.5618 | 0.06206 | - | - | - | 10.0 | 3.0 | Official-initialized full matrix Stage2 spatial diagnostics. |
Official full Stage2 spatial adapterpact_ot_matrix_official_20260521_full_stage2_spatial_adapter | official-full | 40 | 5.9811 | 0.157 | - | - | - | 6.0 | 1.0 | Official-initialized full matrix Stage2 spatial adapter. |