ACTIVE · AWS NVIDIA L4 MI-DETR 1280 Effective-Scale AUG V2 Epochs 10–12 remain active. Best so far is 0.708067 at epoch 6: promising, but currently below the accepted MI-DETR1024 control at 0.715463918.
NEXT · LAB Co-DINO 1280 Effective-Scale AUG V2 Fresh Objects365 initialization after Full-Aug V1 is retired. Prefer RTX5090 at a qualified 450 W limit; use RTX4090 if the device remains meaningfully unstable.
Co-DINO ViT-L + O365 @1280 · variant registry Variant Geometry / recipe Local mAP50 Hidden Status / interpretation Scale-control flip-only · 12 epochs · BS1 × accum8 · FP16 AMP · seed 2026 0.7216258164455305 0.7312659 COMPLETE · ACCEPTED · HIDDEN VERIFIED Full-Aug V1 native-style resize/crop → final canonical 1280 resize/letterbox 0.554091 — FAILED SCIENTIFIC CANDIDATE · SUPERSEDED · shared-epoch deficit approximately -0.159641 Effective-Scale AUG V2 resize / optional crop → fixed 1280 crop/pad → no second scale-normalizing resize — — PLANNED / NEXT · fresh Objects365 initialization · partial loss normalized by actual microbatch count
Co-DINO 1280 learning evidence · Scale-control vs Full-Aug V1 Epoch Scale-control Full-Aug V1 1 0.530777 0.393262 2 0.604291 0.444905 3 0.647740 0.474263 4 0.676348 0.456134 5 0.687198 0.545012 6 0.681855 0.544920 7 0.701045 0.552719 8 0.706751 0.554091 9 0.718312 0.552238
Why Co-DINO Full-Aug V1 failed V1 composed native-style stochastic resize/crop with a final canonical 1280 resize/letterbox. On resize-only paths, the final resize normalized away much of the intended scale variation. It also had a minor partial-window issue: 12,105 training images modulo accumulation8 leaves one residual microbatch that could be normalized as though it contained eight samples. The conclusion is that this composed V1 implementation failed—not that native/full augmentation is universally bad.
MI-DETR Swin-L · variant registry Variant Geometry / precision / batch Local mAP50 Hidden Status / interpretation 1024 scale-control RandomFlip-only + centered letterbox · FP32 · BS1 × accum16 0.7154639176139567 0.7310095 COMPLETE · ACCEPTED · HIDDEN VERIFIED 1280 scale-control clean resolution control approximately 0.5875 at actual epoch 1 / optimizer update 757 — INTENTIONALLY STOPPED / DEPRIORITIZED · NOT FAILED 1280 AUG V1 resize/crop → final canonical 1280 resize — — SUPERSEDED / PRUNED · scale cancellation 1280 Effective-Scale AUG V2 native-inspired resize/crop → fixed 1280 canvas · no second resize best 0.708067 — RUNNING · promising but below 1024 control
MI-DETR 1280 Effective-Scale AUG V2 · active trajectory Actual epoch Canonical mAP50 1 0.575506 2 0.657886 3 0.673305 4 0.699389 5 0.695235 6 0.708067 · current best 7 0.697858 8 0.701916 9 0.701988 10–12 pending / active training
MI-DETR 1280 Effective-Scale AUG V2 · active authority Field State Machine AWS NVIDIA L4 Precision FP16 AMP · FP32 parameters / optimizer state Batch / schedule physical 1 · accumulation16 · effective16 · 12 epochs · 757 successful updates/epoch · 9079 max updates Validation canonical centered 1280 letterbox Reference target accepted MI-DETR1024 = 0.7154639176139567 Decision boundary finish epoch 12 first; then test ensemble replacement only if competitive
MI-DETR next decision Finish V2 through epoch 12. If it becomes competitive with or beats 0.7154639, test it as a replacement inside the current ensemble. If it remains clearly below the 1024 control, the preferred next experiment is a full MI-DETR1280 scale-control—not an automatic 24-epoch V2 retrain. A short terminal-LR continuation is considered only if V2 finishes very close and still improves after late LR decay.
Crop rescorer · proven component, reproduction handoff pending Field Current record Component ConvNeXt-Tiny crop rescorer Local checkpoint checkpoints/safe_crop_rescorer_v1/rescorer_best.pth Checkpoint size / SHA256 111,447,199 bytes / c4af4f50ecc279b5a4e4d6672c0dab6533062064663c930ad4ba7477f1b5d5c0 Status evolution HIGH-PRIORITY EXPERIMENT → PROVEN HIGH-VALUE COMPETITION COMPONENT Evidence teammate ensemble + rescorer reached approximately 0.7701 hidden Caveat exact head, preprocessing, score combination, application recipe, and artifact handoff are still synchronizing
ACTIVE COMPUTE AWS L4 · MI-DETR V2 Epoch 9 completed; epochs 10–12 active. Current best 0.708067.
LAB NEXT Co-DINO Effective-Scale V2 Preferred RTX5090 at 450 W only after bounded sm_120/stability qualification; RTX4090 is the fallback if instability persists.
RETIRED Co-DINO Full-Aug V1 Scientifically failed composed implementation; preserve evidence, but do not continue it.