Marine DebrisCompetition research · v7

EXPERIMENTS · SNAPSHOT 24 AUG · 1 ACTIVE · VARIANT-LEVEL STATUS

Variants & Active Experiments

This page records clean controls, failed implementation compositions, intentionally stopped work, and the one active experiment separately. A low result is never silently relabeled as a model-family failure.

MI-DETR V2 RUNNINGCo-DINO V2 NEXTFULL-AUG V1 FAILEDRESCORER PROVEN
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
VariantGeometry / recipeLocal mAP50HiddenStatus / interpretation
Scale-controlflip-only · 12 epochs · BS1 × accum8 · FP16 AMP · seed 20260.72162581644553050.7312659COMPLETE · ACCEPTED · HIDDEN VERIFIED
Full-Aug V1native-style resize/crop → final canonical 1280 resize/letterbox0.554091FAILED SCIENTIFIC CANDIDATE · SUPERSEDED · shared-epoch deficit approximately -0.159641
Effective-Scale AUG V2resize / optional crop → fixed 1280 crop/pad → no second scale-normalizing resizePLANNED / NEXT · fresh Objects365 initialization · partial loss normalized by actual microbatch count
Co-DINO 1280 learning evidence · Scale-control vs Full-Aug V1
EpochScale-controlFull-Aug V1
10.5307770.393262
20.6042910.444905
30.6477400.474263
40.6763480.456134
50.6871980.545012
60.6818550.544920
70.7010450.552719
80.7067510.554091
90.7183120.552238
MI-DETR Swin-L · variant registry
VariantGeometry / precision / batchLocal mAP50HiddenStatus / interpretation
1024 scale-controlRandomFlip-only + centered letterbox · FP32 · BS1 × accum160.71546391761395670.7310095COMPLETE · ACCEPTED · HIDDEN VERIFIED
1280 scale-controlclean resolution controlapproximately 0.5875 at actual epoch 1 / optimizer update 757INTENTIONALLY STOPPED / DEPRIORITIZED · NOT FAILED
1280 AUG V1resize/crop → final canonical 1280 resizeSUPERSEDED / PRUNED · scale cancellation
1280 Effective-Scale AUG V2native-inspired resize/crop → fixed 1280 canvas · no second resizebest 0.708067RUNNING · promising but below 1024 control
MI-DETR 1280 Effective-Scale AUG V2 · active trajectory
Actual epochCanonical mAP50
10.575506
20.657886
30.673305
40.699389
50.695235
60.708067 · current best
70.697858
80.701916
90.701988
10–12pending / active training
MI-DETR 1280 Effective-Scale AUG V2 · active authority
FieldState
MachineAWS NVIDIA L4
PrecisionFP16 AMP · FP32 parameters / optimizer state
Batch / schedulephysical 1 · accumulation16 · effective16 · 12 epochs · 757 successful updates/epoch · 9079 max updates
Validationcanonical centered 1280 letterbox
Reference targetaccepted MI-DETR1024 = 0.7154639176139567
Decision boundaryfinish epoch 12 first; then test ensemble replacement only if competitive
Crop rescorer · proven component, reproduction handoff pending
FieldCurrent record
ComponentConvNeXt-Tiny crop rescorer
Local checkpointcheckpoints/safe_crop_rescorer_v1/rescorer_best.pth
Checkpoint size / SHA256111,447,199 bytes / c4af4f50ecc279b5a4e4d6672c0dab6533062064663c930ad4ba7477f1b5d5c0
Status evolutionHIGH-PRIORITY EXPERIMENT → PROVEN HIGH-VALUE COMPETITION COMPONENT
Evidenceteammate ensemble + rescorer reached approximately 0.7701 hidden
Caveatexact 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.