feat(mot): add causal routing audit, utility router and adaptive K - #190
Open
blues-kun wants to merge 15 commits into
Open
feat(mot): add causal routing audit, utility router and adaptive K#190blues-kun wants to merge 15 commits into
blues-kun wants to merge 15 commits into
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR:MoT 序列级/目标级路由审计、检测效用监督与 P5 混合架构
关联 Issue #54。
技术总结 Discussion:
#191
改动摘要
P50/P95/P99;
上游已经包含 Issue 指定的三项
tests/test_mot.py边界测试,本分支负责复验,不将其声明为新增。新增测试聚焦 TIFF、dtype、统计、序列配对、遮挡匹配和 benchmark 状态隔离。
与已合并 PR 的差异
和边界测试;本 PR 的结构增量是 P5-only MoT;
并在限制中指出 COCO/VisDrone 使用不同训练域模型;
因此建议将本 PR 作为 #54 的方法学修正与增量结果评审,而不是重复合入已有基础实现。
受控实验
四模型均从 YAML 初始化,在 VisDrone 上训练 30 epoch,640,batch 16,seed 42,FP32。
训练 mAP 来自训练提交
58cb439;当前 mAP 与延迟来自合并上游d5afc4b后的兼容性复验。benchmark 使用同一 RTX 5090、固定输入、每模型至少 2 秒预热、3 轮执行和顺序轮换。
MoT-P5 相比完整 MoT:当前 mAP50-95 增加 0.129 个百分点,P50 降低 45.01%,FLOPs 降低
32.22%。相比 EsMoE:mAP50-95 仅增加 0.099 个百分点,P50 增加 28.03%。因此只声明它是
完整 MoT 的低预算替代,不声明相对 EsMoE 已产生协同。
路由结论
+0.001575,CI 跨 0,q=0.703;+0.000674,仅作 OOD 信号;top_k/exploration/straight-through,不在本 PR 中无实验修改默认行为。更细的目标级匹配审计保留 12,296 对、76 个序列。
model.14.m.0中高遮挡目标的LocalConv/Deformable 框内概率分别增加 0.00693/0.00259,Window 降低 0.00952;这说明是
组合重分配,不支持硬编码“遮挡切 Deformable”。
检测效用路由结果
model.14.m.0;36.9%/35.7%/27.4%;
但 test-dev 升至 0.05476;
0.08743 → 0.08695,P5026.877 → 28.244 ms;验证
2026-07-29 在 Python 3.9.25、PyTorch 2.8.0+cu128 上复验:
134 passed, 18 warnings。其中 14 个 warning 是热力图测试触发的 Matplotlib/pyparsing 弃用提示,4 个来自既有 MoA head
数量自动调整;无失败。复验同时修复了 Python 3.8/3.9 不支持
typing.Self的导入问题。RTX 5090 CUDA smoke 能加载同一 checkpoint 与 utility bundle,目标层选出
K=1,上下文退出后恢复原固定-K 配置。
范围与限制