dfine-large-fire-smoke-v7

Documentation publique

An experimental two-class D-FINE Large detector fine-tuned with DEIM for FireViewer. This repository contains the evaluated V7 EMA weights, a self-contained native inference runtime, and a transparent external benchmark against two frozen reference models.

Status: experimental, not production/emergency ready. On the selected 212-image Home-Fire holdout: mAP50:95 0.2859, AP50 0.4996, AP75 0.2798, precision 0.7178, recall 0.4349, F1 0.5417. Recall fails the 0.50 minimum; the corpus lacks enough negative images and verified independent events for a support-grade claim. The benchmark's execution passed, its support-grade admission did not.

What is included

no optimizer, scaler, pickle or training dataset.

CSV, source/selection hashes, overlap evidence and scoring-source snapshots.

Publication checks: all 1,253 exported tensors are exactly equal to the source EMA state. Reloaded BF16 predictions match the original benchmark adapter exactly on three in-memory synthetic probes, including filtered outputs at 0.05 and 0.5. See export verification. This proves export/reload equivalence, not additional detection quality on real images.

• model.safetensors: 125.14 MB, exact EMA tensors from the evaluated checkpoint;
• inference.yml, fireviewerinference.py, and the pinned fireviewerdeim/ runtime.
• Benchmark report, all 12 results (JSON),
• Training summary and component/source-rights notices. Corpus images and annotations are not redistributed.

Run inference

Tested with Python 3.11.9, torch 2.11.0 / torchvision 0.26.0, CUDA 12.8, and RTX 5070 Ti. Choose a PyTorch build matching your GPU/driver first. Then:

For immutable deployment, add --revision to hf download. CPU fallback: --device cpu --precision fp32; CPU speed/quality was not benchmarked. The helper uses Safetensors and verifies the model/config hashes before loading. No pretrained backbone download, training setup, private path, HF token, or additional FireViewer repository is required. This is a native DEIM checkpoint/runtime, not a Transformers AutoModel-compatible export.

RGB is directly resized to 704x704, converted to [0,1] tensors, without letterboxing or mean/std normalization. FP32 parameters use BF16 autocast on CUDA. Output is score-sorted, at most 100 boxes by default, with confidence >=0.5. Box centers are not validated fire-origin points or geolocations.

Training provenance

V7 selected corpus: 5000 images (4002 train / 486 validation / 512 internal test), 40 total epochs, 704x704, seed 42. Resumed stage: micro-batch 2, accumulation 8, effective batch 16. This release uses beststg2.pth, EMA branch, not the final/optimizer state. Checkpoint SHA-256: 7b26fbf4cf05723670c1fd01da4542ca98fb92aef1e2f0c14dfde4f86184dec5. Safetensors SHA-256: a43d2fa1269e770e4dd5fa303d2f36591c1d9cb5a8351880c8e6b05c837cd136.

Base: DEIM D-FINE HGNetv2-L COCO 50e. See training summary and runtime provenance. No new training was performed for publication.

Evaluation and limits

Same selected images/annotations, operating confidence 0.5, IoU 0.5. AP uses candidate confidence 0.05. V7 is 704 BF16; references are 640 FP16, so this is not a controlled architecture comparison. Source independence is limited to declared training sources plus a complete V7 exact/pHash train-overlap screen. The two comparators' full training media were not independently screened.

Only 12 negative images and no verified independent event IDs are available. The sample is selected, not a comprehensive field validation. JPEG50 produces 3/12 negative false alarms. Full definitions, per-class AP, robustness, precision/recall, threshold diagnostics and measurement limits are in the report. Latency was measured on one image, not as end-to-end full-corpus throughput.

License, attribution and safety

Read LICENSENOTICE.md before reuse. Code licensing is separate from fine-tuned weights and source data rights; public availability is not a blanket commercial-use grant. Credit the DEIM authors, D-FINE/RT-DETR contributors, and Home-Fire Dataset for the evaluation source. Do not use as the sole basis for emergency alerts, autonomous firefighting or safety-critical decisions. No segmentation, physical ignition-point, geolocation, or event-level field-performance claim is made.