--- license: bsd-3-clause tags: - fireviewer - convnext - torchvision - fire - smoke - image-classification library_name: pytorch pipeline_tag: image-classification datasets: - fireviewer/fireviewer_fire_smoke_flame_v3_native model-index: - name: FireViewer ConvNeXt-Tiny · Context results: - task: type: image-classification name: Image Classification dataset: type: fireviewer/fireviewer_fire_smoke_flame_v3_native name: FireViewer · labelled validation subset config: default split: validation revision: 8c8b197fa1c928217529a04bff434fb70a66614e metrics: - type: f1 name: F1 macro (%) value: 76.92 - type: accuracy name: Accuracy (%) value: 87.72 --- ![FireViewer · ConvNeXt-Tiny · Six visual contexts](assets/card-header.svg) # ConvNeXt-Tiny · Six visual contexts An image-level classifier with **six context outputs**. Targets come from the Luna context annotations of the FireViewer corpus. It predicts one category for the whole image, without bounding boxes. | Input | Validation macro F1 | Validation accuracy | |---|---:|---:| | **920 × 920 RGB** | **76.92%** | **87.72%** | ## Checkpoint and class order [`best.pth`](best.pth) is the selected validation checkpoint. [architecture.json](architecture.json) records the Torchvision architecture and initialization; [class_mapping.json](class_mapping.json) fixes the output order. | ID | Category | Interpretation | |---:|---|---| | 0 | `human_useful_fire` | Controlled human use | | 1 | `domestic_origin` | Domestic context | | 2 | `suspect_origin` | Undetermined or insufficient origin detail | | 3 | `fire_road_urban` | Road / urban | | 4 | `forest_natural_area_wildfire` | Apparent uncontrolled vegetation fire | | 5 | `negative` | No visible fire or smoke | These are visual-context categories, **not an ordinal severity scale**. `suspect_origin` does not establish criminal intent; it covers insufficient origin information or detail with potentially hazardous fire or smoke. One image can contain multiple contexts even though the model returns one class. ## Validation results ![Class F1, macro F1 and accuracy on the labelled validation subset](assets/validation-summary.png) The best checkpoint is evaluated at **epoch 5 on 11,237 validation images with a context label**. Prediction is the highest of six logits. | Category | Precision | Recall | F1 | Images | |---|---:|---:|---:|---:| | `human_useful_fire` | 62.39% | 78.46% | 69.51% | 260 | | `domestic_origin` | 47.46% | 58.33% | 52.34% | 48 | | `suspect_origin` | 74.71% | 75.26% | 74.98% | 2,292 | | `fire_road_urban` | 88.35% | 84.57% | 86.42% | 1,121 | | `forest_natural_area_wildfire` | 80.20% | 87.47% | 83.68% | 1,181 | | `negative` | 95.76% | 93.43% | 94.58% | 6,335 | Macro F1 gives equal weight to all six classes; accuracy is more influenced by frequent classes. The domestic class has only **48 validation images**, so a few errors can materially change its score. These values measure agreement with Luna labels, not independent real-world danger assessment or held-out test performance. ## PyTorch inference ```bash pip install torch torchvision pillow huggingface_hub ``` ```python import torch from PIL import Image from torch import nn from torchvision.models import convnext_tiny from torchvision.transforms import functional as TF from huggingface_hub import hf_hub_download path = hf_hub_download( "fireviewer/convnext-tiny-alert-categories-v2", "best.pth", ) checkpoint = torch.load(path, map_location="cpu", weights_only=False) labels = checkpoint["classes"] size = int(checkpoint["image_size"]) model = convnext_tiny(weights=None) model.classifier[2] = nn.Linear(model.classifier[2].in_features, len(labels)) model.load_state_dict(checkpoint["model"]) model.eval() with Image.open("image.jpg") as image: image.draft("RGB", (size, size)) image = image.convert("RGB").resize((size, size), Image.Resampling.BILINEAR) tensor = TF.normalize( TF.to_tensor(image), mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], ).unsqueeze(0) with torch.inference_mode(): logits = model(tensor) category = labels[int(logits.argmax(dim=1).item())] print(category) ``` Preprocessing is RGB, direct bilinear resizing to **920 × 920**, pixels in `[0,1]`, followed by ImageNet normalization: mean `[0.485,0.456,0.406]`, standard deviation `[0.229,0.224,0.225]`. It does not use YOLO letterboxing. ## Use, rights and limitations Use the model to explore, filter or group images by visual context. It does not locate objects, estimate camera-to-fire distance or determine the historical cause of an incident. Ambiguous scenes and camera, illumination or environment shifts can reduce performance. Declared licence: **BSD-3-Clause**. Image rights are described separately in the [dataset card](https://huggingface.co/datasets/fireviewer/fireviewer_fire_smoke_flame_v3_native). Evaluated weights originate from revision `12dd4da71da5dd4264c63e07cab5f41c3a007b77`; the earlier package identity remains in its provenance. This English presentation does not change the checkpoint. [FireViewer](https://huggingface.co/fireviewer) · [Resource catalogue](https://huggingface.co/spaces/fireviewer/README/blob/main/docs/RESOURCE_INDEX.md) · Presentation updated **14 November 2026**.