Spaces:
Configuration error
Configuration error
|
Download docs/CHOOSING_VARIANTS.md from fireviewer/README: direct link, hf CLI and curl.
- Browser
- Download file 3.41 kB
-
https://huggingface.co/spaces/fireviewer/README/resolve/main/docs/CHOOSING_VARIANTS.md
- Command line
-
hf download hf://spaces/fireviewer/README/docs/CHOOSING_VARIANTS.md
-
curl -L -o CHOOSING_VARIANTS.md https://huggingface.co/spaces/fireviewer/README/resolve/main/docs/CHOOSING_VARIANTS.md
3.41 kB
| # Choose a FireViewer resource | |
| [Home](../README.md) · [Full catalogue](RESOURCE_INDEX.md) | |
| | Need | Resource | Contract to preserve | | |
| |---|---|---| | |
| | Locate visible flame and smoke with the retained YOLO11M | [YOLO11M A100](https://huggingface.co/fireviewer/yolo11m-flame-smoke-resume-a100) | RGB, 928 px letterbox; 0=flame, 1=smoke. Best checkpoint and latest measurement are separate. | | |
| | Examine a smaller-checkpoint detector | [YOLO26s L4](https://huggingface.co/fireviewer/yolo26s-flame-smoke-l4) | Distinct architecture; 928 px, 0=flame, 1=smoke. Lower recorded AP; device performance is not measured here. | | |
| | Classify an image’s visual context | [ConvNeXt-Tiny](https://huggingface.co/fireviewer/convnext-tiny-alert-categories-v2) | Direct RGB 920×920 resize and ImageNet normalization; six non-ordinal contexts, no boxes. | | |
| | Use recent Cadryl conversions | [YOLO11M](https://huggingface.co/unicornwhodev/lite-rt_models_cadrylpro/tree/main/models/fireviewer_yolo11m_fire_smoke_928_trainable) / [ConvNeXt](https://huggingface.co/unicornwhodev/lite-rt_models_cadrylpro/tree/main/models/fireviewer_convnext_tiny_alerts_920_trainable) | Read the central package signatures, exact tensors, preprocessing and runtime. Colab checks are not physical-device qualification. | | |
| | Recover the historical conversion source | [YOLO11M E42](https://huggingface.co/fireviewer/yolo11m-fire-smoke) | 0=fire, 1=smoke; earlier training snapshot and distinct weights, retained for source identity. | | |
| | Inspect older conversions | [LiteRT archive](https://huggingface.co/fireviewer/litert-models) | Read each variant contract. Evidence for one variant does not qualify the others. | | |
| | Train a two-class detector | [Fire & Smoke v3](https://huggingface.co/datasets/fireviewer/fireviewer_fire_smoke_flame_v3_native) | Native-pixel XYWH boxes; 0=flame, 1=smoke; preserve splits and source authority. | | |
| | Propose fire-or-smoke regions under one class | [Probable fire v3](https://huggingface.co/datasets/fireviewer/fireviewer_probable_fire_v3_native) | 0=probable_fire; distinct corpus and population. | | |
| ## Compare the right evidence | |
| Pin the model and dataset SHA, preprocessing, evaluated population and split. A selected best score and a final snapshot may describe different checkpoints. Classification F1, detector AP and cascade diagnostics are different measurements. | |
| Label order differs by release: the historical RT-DETR uses **0=smoke_visible, 1=flame_visible**, unlike the recent YOLO detectors. Read the matching configuration rather than assuming a universal mapping. | |
| The completed seven-human-lot continuation did not improve the retained YOLO11M checkpoint and produced the same best-weight hash. Its duplicate repository was retired; [the comparison receipt](https://huggingface.co/fireviewer/yolo11m-flame-smoke-resume-a100/blob/main/evaluation/retired-seven-lot-run.json) retains the evidence. | |
| ## Historical material and use boundaries | |
| The [v2.5 cascade](CASCADE_V25.md) and [v2.5 dataset guide](DATASET_V25.md) retain their original contracts and limitations. They are historical references rather than the current model entry point. | |
| Visual context does not determine actual danger. Published weights and technical checks do not establish human acceptance, physical-phone qualification or operational emergency performance. Source-specific image, checkpoint and software terms remain applicable. | |