CNN-based Persian Digit Recognizer (PyTorch)

Role

Developer - AI/ML Project

Project Summary

This project delivers a full Persian digit recognition pipeline (0..9) with a custom PyTorch CNN and reproducible CLI workflows. I implemented dataset preparation, model training, evaluation, and single-image inference scripts in one clean project structure.

Repository

CNN-based-Persian-digit-recognizer-built-with-PyTorch

What We Built

• Custom model PersianDigitCNN for 10-class classification.
• HODA dataset preparation pipeline with automated output generation.
• Training workflow with checkpointing and run metrics.
• Evaluation workflow on validation/test splits.
• Prediction script for single-image inference (top-k output).
• Confusion matrix generation for validation and test analysis.

Model & Training Setup

• Architecture:
  • Conv(1->32) + BatchNorm + ReLU + MaxPool
  • Conv(32->64) + BatchNorm + ReLU + MaxPool
  • Conv(64->128) + BatchNorm + ReLU + MaxPool
  • AdaptiveAvgPool(3x3)
  • Linear(1152->256) + ReLU + Dropout(0.3) + Linear(256->10)
• Loss: CrossEntropyLoss
• Optimizer: AdamW
• Dataset: HODA

Results

• Best validation accuracy: 99.76%
• Test accuracy: 99.56%
• Test loss: 0.0169

Confusion Matrices

Skills

• AI
• Python
• Deep Learning
• PyTorch
• Computer Vision
• CNN Architecture
• Model Evaluation