16.1 The winning solution
A detailed description of the winning solution is given at https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/301686
The team consisted of:
- Giba
- Competition grandmaster
Giba`s description is as follows:
Giba`s description is as follows:
So basically the challenge is a regression and we were asked to predict the Pawpularity of each pet given the image and some tabular meta features.
The solution consists of two parts
- PART 1. Transfer Learning + SVR
- The idea was to extract features from pretrained architectures to transform the problem into a tabular approach, then fit a model using that data. To extract features from images he used imagenet pretrained models available in timm and OpenAI CLIP libraries.
- PART 2. CNN + Vision Transformers ensembling:
- This part of solution is an weighted average of 5 classical image regressions models using different backbones and augmentations.