This challenge is aimed on creation of AI that capable to determine a genre of an artwork based on its visual appearance. Visual appearance of an artwork is represented by photos that are carefully distributed into categories.
This was a wonderful opportunity to try and implement error correction code approach to classification. Rather than creating a number of separate binary classifiers, the neural network was designed to be trained on class codes rather than plain classes.
An elegant non-iterative self calibrating solution to probability estimates for such a classifier from Reducing multiclass to binary by coupling probability estimates by Bianca Zadrozny was used.
This solution scored 4th place on a public test and 12th place on a private test at https://codenrock.com/contests/masters-of-arts#/
Units required:
- Pandas
- PyTorch
- skorch
- timm
- scikit-learn
- PIL
- NumPy
- Matplotlib
- Seaborn