CURRENT · TRAINING — RUNNING

DEIM-D-FINE-X + Objects365 — Production Training

Canonical DEIM-D-FINE-X production run using HGNetv2-B5 and Objects365 initialization on the primary RTX4090 competition path.

TRAINING — RUNNINGPRODUCTION PREFLIGHT · PASSPRIMARY GPU · RTX4090
RUNNINGproduction training
24epochs
BS4physical batch
1024 × 1024resolution
Canonical production configuration
FieldSetting
ModelDEIM-D-FINE-X
BackboneHGNetv2-B5
InitializationObjects365
Classes34
Resolution1024 × 1024
Physical batch4
Gradient accumulationNone
Seed2026
Epochs24
Checkpoint cadenceEvery epoch
GPURTX4090
warmup_iter2000
Fresh integration and production gates
GateStatus
Gate APASS
Gate BPASS
GPU calibrationACCEPTED WITH EVIDENCE LIMITATION
Production preflightPASS
Production trainingRUNNING
RTX4090 calibration evidence
Physical batchThroughputInterpretation
15.748 images/sMeasured
27.950 images/sMeasured
49.479 images/sSelected production batch
8OOM observedSupporting evidence only; raw record was not serialized

Canonical project checkpoint selection after training

24 retained epoch checkpoints
Standalone evaluationUse the existing project AP50 evaluator.
Highest reproducible project mAP50If tied, select the earliest epoch.
Canonical DEIM best checkpoint

Reproducibility strategy

  • frozen dataset split
  • pinned upstream revisions
  • checkpoint SHA256
  • qualified per-model environment
  • deterministic dataset adapter
  • fixed seed 2026

Acceptance evidence

  • standalone project metric evaluation
  • best reproducible project mAP50
  • evidence freeze after the accepted run