EDA and training procedure for multitask age and gender classification
I used OUI-Adience dataset and trained Resnet50 and MobilenetV3 models
Final accuracy: gender 96.94%, age 78.79% (resnet50), gender acc 96.12%, age acc 82.12% (mobilenet_v3)
MobilenetV3 implementation has taken from pytorch-image-models
Pretrained models:
- resnet50: to-be-published
- mobilenetv3: link
Model expects 224x224 image and outputs age_logits (bs x 8), gender_logits (bs x 3) age logits - probability of image for each of 8 groups: 0-2, 3-6, 8-13, 15-20, 22-32, 34-43, 45-53, 55-100