Bangla Grapheme Prediction

Why grapheme recognition is a multi-label problem

Bengali handwriting doesn't reduce to single-character classification: each glyph is composed of a grapheme root, an optional vowel diacritic, and an optional consonant diacritic, combined into thousands of visually distinct graphemes from only 168 root, 11 vowel, and 7 consonant components. The Kaggle competition (Bengali.AI Handwritten Grapheme Classification) scores all three independently, so the model has to predict three labels per image, not one.

Model and pipeline

• A shared SEResNeXt50 (se_resnext50_32x4d) backbone with three linear classification heads, one per label, trained jointly.
• Data loaded from feather files instead of the source Parquet: roughly 2s per shard vs. ~60s. That's the difference between iterating on augmentation choices in an afternoon vs. a day.
• Augmentation stack: affine transforms (rotation, scale, shear, translation) plus albumentations (grid distortion, cutout, piecewise affine, brightness/contrast, blur and Gaussian noise), needed because the same grapheme is drawn with real variation in stroke weight and slant across writers.
• Training loop built on pytorch-ignite (Engine/Events), Adam at lr 1e-3 with ReduceLROnPlateau, over 100 epochs.
• Metric: macro recall per head, combined as a weighted average (root weighted 2x vowel and consonant, matching the competition's own scoring) rather than plain accuracy. Root prediction is both the harder problem (168-way) and the one the score cares about most.

Backbone and training-loop scaffolding adapted from the public SEResNeXt reference kernel for this competition; the multi-head classifier, augmentation tuning, and training runs are this submission's work.