---
language: en
tags:
  - pokemon
  - card-grading
  - image-classification
  - onnx
metrics:
  - accuracy
---

# Poke-Grader Defect Classifier

Fine-tune de `microsoft/swin-small-patch4-window7-224` para clasificar defectos en cartas Pokémon TCG.

## Cabezas de clasificación

| Cabeza | Clases |
|---|---|
| `surface_scratch` | none / micro / light / medium / heavy |
| `corner_wear` | none / light / medium / heavy |
| `edge_whitening` | none / light / medium / heavy |

## Métricas (mejor epoch: 28)

| Métrica | Valor |
|---|---|
| val_loss | 0.1163 |
| Exactitud surface | **93.3%** |
| Exactitud corners | **96.0%** |
| Exactitud edges | **99.8%** |
| **Exactitud media** | **96.4%** |

Entrenado con **5000 cartas** (~mitad inglés, ~mitad japonés) durante **30 épocas** de fine-tuning.

## Uso (ONNX)

```python
import onnxruntime as ort
import numpy as np

sess = ort.InferenceSession('card_defect_classifier.onnx')
# img: np.array float32, shape (1, 3, 224, 224), normalizado ImageNet
surface, corners, edges = sess.run(None, {'pixel_values': img})
```

Clases: `['none','micro','light','medium','heavy']` / `['none','light','medium','heavy']` / `['none','light','medium','heavy']`