Multi-Script Router — T3.7
A compact CNN (~330k parameters) that classifies handwritten characters into four Indian scripts: Devanagari, Tamil, Bengali, and Kannada.
Architecture
- 3 × ConvBlock (residual skip connections)
- Global Average Pooling → FC 128 → 256 → 4
- Input: 32×32 grayscale
Training
| Metric | Value |
|---|---|
| Best val acc | 78.36% |
| Test acc | 83.16% |
| Epochs | 15 |
| Batch size | 128 |
| Optimizer | AdamW + CosineAnnealing |
W&B Run
Training logged at: https://wandb.ai/models-iiit-hyderabad6183/mnist