C3 vpeft smoke - #265
Closed
lycyhrc wants to merge 10 commits into
Closed
Conversation
…/10/50/100, env manifest - conda env(py3.11)@user workspace: torch 2.6.0+cu124, yolo 8.4.101, cuda on 8x A800 - NEU-DET: 1800 imgs/6cls, train 1620/val 180, shots k5=30 k10=60 k50=300 k100=600, MIT-unknown license noted - DeepPCB: 1500 imgs(640x640 jpg)/6cls, train 1350/val 150 seed824, class-imbalance few-shot k5=8 k10=13 k50=70 k100=142 (use NEU-DET for the P2 curve) - datasets/manifest.md: sources, licenses, sha256 samples, reproducibility notes
…01 bare-yolo PATH pollution - server_run.py: path-inferred runner (tag, refuse-overwrite, gpu snapshot, training.log, summary.json) - smoke results (GPU5, NEU-DET full 1620): vpeft(2ep,strict planner ok) / full_sft(1ep) / frozen_backbone(1ep, mem 1.37G vs full 4.54G) - planner_solver_notes.md: source-level facts (solvers ao/dco/mip, OR-Tools ImportError, cap<8 at placement_plan.py:117, audit/LOVO paths) - bugfix BF-01: bare 'yolo' on shared box resolves to wrong ultralytics (chenzhong1 repo) -> use Path(sys.executable).parent/yolo in runner & smoke/c3/run_smoke.py - stage3 decision points: warmup_epochs=3.0 vs epochs budget; EsMoE-N full params only 2.8M -> model size for P1 param-efficiency argument TBD
…-race), server_run --lora-budget
- make_matrix.py: {neu,pcb} x {vpeft,full_sft,frozen_backbone} x seed{824,2024,777}, epochs=100
- sweep_runner.py: idle-GPU scheduling, matrix-state breakpoint resume, orphan probing; subprocess env strips PYTHONPATH (CodeBuddy sitecustomize shim wraps unlink -> OSError)
- collect_evidence.py: per-unit 4-dim evidence (mAP/elapsed/gpu_mem/resolved config + [V-PEFT] lines)
- BF-02 documented: concurrent labels.cache first-build race (ultralytics data/utils.py:891 unlink) amplified by env shim; neu_vpeft_s824 retried as s824b
…utlier diagnosis + supplement runner - results: NEU full_sft ~0.75-0.78 vs vpeft 0.69-0.75 (s2024 stuck 0.37) vs frozen 0.66-0.69 (s2024 dead 0.21); PCB full 0.989-0.99 > vpeft 0.82-0.96 > frozen 0.64-0.91 (s824 oscillating) - measured trainable: full 2,813,626 / vpeft LoRA adapter 116,736 (+unfrozen class-mismatch head 348,514) - peak mem: full 5.3G > vpeft 4.15G > frozen 3.7G - collect_evidence v2: mAP=best over epochs (last-row biased: frozen s2024 last 0.023 vs best 0.210), peak GPU parsed from training.log GpuMem - run_supplement.py: seed-2025 NEU vpeft/frozen + PCB frozen s824 rerun to confirm determinism of outliers
…ble), planner audit summary, README; finalize after supplement runs
…inism verdict) - stage3 README: supplement results table (neu_vpeft_s2025=0.644 OK, neu_frozen_s2025=0.739 OK => s2024 anomalies were seed-specific; pcb_frozen_s824b=0.639@ep9 identical rerun => oscillation deterministic) - stage4 README finalized: full 0.767+-0.014 (neu) / 0.989+-0.001 (pcb); vpeft robust-subset 0.694; paired diff full-vpeft +0.165(neu)/+0.108(pcb); 4-dim table - analysis_stats: supplement discovery + 824b labeling; planner_audit summary; p2/ planner-bug issue draft (placement_plan.py:113-118 ValueError when rank>cap) + case-embedding note - stage5: report_final.md filled; reproduction/ assembled; README status update
…ework - assertion headlines per slide (problem→protocol→3 findings→decision→evidence→limits) - visual: dark cover/close, teal scan motif, big-number stats, hand-drawn bars & seed dot-spread, per-slide evidence footer - auto QA: layout overflow + textbox overlap + content-completeness checks pass
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
2 participants
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.
No description provided.