pawpularity-kaggle

Platform for Pawpularity Kaggle contest https://www.kaggle.com/c/petfinder-pawpularity-score

Models tried

• Gamma Regression on meta data: RMSE ~20
• Mixture Model on meta data, distributions adopted are Gamma for pawpularity and Bernoulli for meta data: RMSE ~20
• Pretrained ResNet18 followed by Gamma Regression: RMSE ~19
• Pretrained ResNet18 followed by MLP: RMSE ~19
• Pretrained ResNext101 followed by MLP: RMSE ~19
• Pretrained ResNext101 followed by Multitask MLP: RMSE ~19

Outcomes

• Multitask improve very little
• Simply Concatenate the meta data improve very little

Further explorations

• Multitask Learning
• Ways to use meta data
• EfficientNet
• Swin Transformer