gender4-whole-image-5class

MobileNetV2 whole-image classifier. Five mutually exclusive classes:

Index	Label
0	real_male
1	real_female
2	anime_male
3	anime_female
4	other

other covers non-person content (objects, animals, scenery).

Files

• model.onnx — weights
• labels.json — label order
• metrics.json — held-out validation metrics

Inference

• Input input: float32 [1, 3, 256, 256], RGB CHW, values in [0, 1] (divide by 255). ImageNet mean/std is inside the graph.
• Output output: float32 [1, 5], softmax probabilities in label order above.

Validation

Accuracy 96.1% (1200 images/class). other → person false-positive rate 0%.