Given the image of a handwritten Bengali grapheme, the challenge is to separately classify three constituent elements in the image: grapheme root, vowel diacritics, and consonant diacritics.
- Used a single Resnet18 Model backbone for all three labels, and used CrossEntropy Loss.
- Grapheme_root loss was given more weight, since the number of classes of grapheme_root were high when compared to the other two.
- 5 - fold stratified split training.