RUNNINGproduction training
24epochs
BS4physical batch
1024 × 1024resolution
| Field | Setting |
|---|---|
| Model | DEIM-D-FINE-X |
| Backbone | HGNetv2-B5 |
| Initialization | Objects365 |
| Classes | 34 |
| Resolution | 1024 × 1024 |
| Physical batch | 4 |
| Gradient accumulation | None |
| Seed | 2026 |
| Epochs | 24 |
| Checkpoint cadence | Every epoch |
| GPU | RTX4090 |
| warmup_iter | 2000 |
| Gate | Status |
|---|---|
| Gate A | PASS |
| Gate B | PASS |
| GPU calibration | ACCEPTED WITH EVIDENCE LIMITATION |
| Production preflight | PASS |
| Production training | RUNNING |
| Physical batch | Throughput | Interpretation |
|---|---|---|
| 1 | 5.748 images/s | Measured |
| 2 | 7.950 images/s | Measured |
| 4 | 9.479 images/s | Selected production batch |
| 8 | OOM observed | Supporting 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