File size: 735 Bytes
eb1d853
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
abb476f
 
eb1d853
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
---
language: [hi, ta, bn, kn]
tags: [image-classification, handwriting-recognition, indian-scripts, pytorch]
license: mit
---

# 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>