File size: 1,722 Bytes
b1ee1c0
7c3eead
 
 
 
b1ee1c0
7c3eead
 
b1ee1c0
 
7c3eead
0e21869
7c3eead
 
 
 
 
 
 
 
 
 
b1ee1c0
 
7c3eead
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
---
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.