Spaces:
Running
Running
| title: DermNet Skin23 Classifier | |
| emoji: 🩺 | |
| colorFrom: blue | |
| colorTo: purple | |
| sdk: gradio | |
| sdk_version: 5.29.0 | |
| python_version: "3.11" | |
| app_file: app.py | |
| pinned: false | |
| license: apache-2.0 | |
| short_description: Skin disease classifier — EVA-02-L, 81.5% accuracy | |
| models: | |
| - iamcode6/dermnet-skin23-eva02 | |
| - iamcode6/dermnet-skin23-convnext | |
| - iamcode6/dermnet-skin23-dinov2g | |
| tags: | |
| - medical | |
| - dermatology | |
| - image-classification | |
| - eva02 | |
| - vision-transformer | |
| # DermNet-Skin23 Classifier | |
| Single-model demo of a 23-class clinical skin disease classifier built on EVA-02-L (~304M params, ViT-L/14) and fine-tuned on a consolidated DermNet + Skin40 dataset. | |
| ## Numbers | |
| | Setup | Accuracy | Macro F1 | | |
| |---|---|---| | |
| | **This Space (single EVA-02-L)** | **81.48%** | **0.7969** | | |
| | Full 5-model ensemble (EVA-02 × ConvNeXt-V1-XL) | 82.86% | 0.8113 | | |
| The full ensemble lives in the linked model repos and takes ~10x more compute — this Space runs the strongest single model, which is good enough for an interactive demo. | |
| ## Dataset | |
| 23 broad dermatology categories merged from DermNet + Skin40 — 17,557 training images, 3,856 validation images. Three small-class stragglers (Stasis_Edema, Stasis_Ulcer, Ichthyosis at 60 images each) were merged into larger neighbors. | |
| ## Training stack | |
| - AMD Instinct MI300X (192 GB HBM3), ROCm 7.0, PyTorch with HIP | |
| - Two-stage fine-tune: 30 epochs at peak LR + 15-epoch continuation at 0.1× LR with mixup off | |
| - bf16 autocast, channels-last memory format, EMA + SWA, weighted-effective-number sampler | |
| - HAM10000 domain pretraining as warm start | |
| ## Disclaimer | |
| For research and educational use only. NOT a diagnostic tool. Always consult a qualified dermatologist for medical concerns. | |