HoneyDog — MobileNetV4-Medium 31-Attribute Facial Predictor

A compact multi-task facial attribute model: one forward pass returns 3 mutually exclusive softmax groups (age, race, face shape) plus 22 binary attributes.

Backbone: MobileNetV4-Conv-Medium (mobilenetv4_conv_medium, 9.11 M nominal parameters), initialised from ImageNet-1K weights and fine-tuned end-to-end on the full VGGFace2 (3,057,398 face images / 8,414 identities).

This is not a transformers model. It is a plain PyTorch checkpoint loaded by the inference.py shipped in this repo — Hub's autogenerated pipeline() snippets do not apply.

Model specifications

Output heads:

• age — Young, Middle_Aged, Senior, OTHER
• race — Asian, White, Black, OTHER
• face_shape — Oval_Face, Square_Face, Round_Face, OTHER
• 22 binary attributes (sigmoid, threshold 0.5): Beard, Eyeglasses, Obstructed_Forehead, Arched_Eyebrows, Male, Rosy_Cheeks, Wavy_Hair, Receding_Hairline, Bangs, Double_Chin, High_Cheekbones, Chubby, Brown_Eyes, Bags_Under_Eyes, Mouth_Closed, Smiling, Big_Nose, Heavy_Makeup, Wearing_Hat, Wearing_Earrings, Wearing_Necktie, Wearing_Lipstick

Evaluation

Held-out, fully frozen test split of 25,401 images (never seen during training). Numbers below are copied from metrics.json in this repo.

Usage

pip install torch torchvision timm pillow

import json
from inference import load, predict

device = "cuda" if torch.cuda.is_available() else "cpu"
model = load(device=device)                      # loads ./model.pt next to inference.py

result = predict(model, "face_example.jpg", device=device)
print(json.dumps(result, indent=2))

result["groups"][<age|race|face_shape>] holds the predicted class, its confidence and the full probability vector; result["attributes"][<attribute>] holds {"present": bool, "confidence": float}.

Command line:

python inference.py face_example.jpg

Files

Limitations

• Attribute labels come from dataset annotation (VGGFace2 images + MAAD-Face annotations); they are coarse, partly noisy, and the OTHER class absorbs anything outside the three named classes.
• VGGFace2 is celebrity-skewed, so age/race/face-shape predictions inherit that bias.
• Accuracy drops on occluded, very low-resolution, or strongly rotated faces; the model was trained on 224×224 crops.
• Single-face, single-image classifier: no detection, no alignment, no identity/recognition.
• Demographic attributes are sensitive. Treat the outputs as noisy estimates for research or dataset-analytics use — not as grounds for decisions about individuals.

Data sources and citations

Trained on VGGFace2 images with MAAD-Face annotations. If you use this model or derivatives, follow the respective dataset licences and cite:

@inproceedings{cao2018vggface2,
  title={VGGFace2: A dataset for recognising faces across pose and age},
  author={Cao, Qiong and Shen, Li and Xie, Weidi and Parkhi, Omkar M and Zisserman, Andrew},
  booktitle={2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018)},
  pages={67--74},
  year={2018},
  organization={IEEE}
}

@article{terhorst2021maad,
  title={MAAD-Face: A massively annotated attribute dataset for face recognition},
  author={Terhorst, Philipp and Huber, Marco and Luu, Nhan and Fayyad, Fadi and Kirchbuchner, Florian and Kuijper, Arjan},
  journal={IEEE Transactions on Information Forensics and Security},
  volume={16},
  pages={3942--3957},
  year={2021},
  publisher={IEEE}
}

@article{qin2024mobilenetv4,
  title={MobileNetV4--Universal Models for the Mobile Ecosystem},
  author={Qin, Danfeng and Leichner, Chas and Delakis, Manolis and Fornoni, Marco and Luo, Shuyang and Yang, Fan and Wang, Weijun and Banbury, Colby and Cheng, Yu and Bapst, Victor and others},
  journal={arXiv preprint arXiv:2404.10518},
  year={2024}
}