https://github.com/parvvaresh/handwritten-persian-numeral-recognition

https://github.com/parvvaresh/handwritten-persian-numeral-recognition

cnn hoda-dataset persian-handwriting pytorch

Last synced: 4 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/parvvaresh/handwritten-persian-numeral-recognition
• Owner: parvvaresh
• Created: 2023-10-28T21:07:05.000Z (about 2 years ago)
• Default Branch: main
• Last Pushed: 2025-03-28T10:26:14.000Z (7 months ago)
• Last Synced: 2025-04-23T12:49:47.997Z (7 months ago)
• Topics: cnn, hoda-dataset, persian-handwriting, pytorch
• Language: Jupyter Notebook
• Homepage:
• Size: 110 KB
• Stars: 4
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Hoda Dataset Reader & CNN Digit Recognizer

This Google Colab notebook demonstrates how to read and preprocess the Hoda handwritten digits dataset, train a Convolutional Neural Network (CNN) using PyTorch, and evaluate its performance. A trained model is saved at the end of the notebook.

## Features

- **Data Loading & Preprocessing**  

  Reads the Hoda dataset (CDB files), resizes, normalizes, and binarizes the images.

- **Data Visualization**  

  Displays sample images with their corresponding labels.

- **CNN Model**  

  Defines a CNN with two convolutional blocks (each with batch normalization, pooling, and dropout) followed by two fully connected layers.

- **Training & Evaluation**  

  - The model was trained for 15 epochs with the Adam optimizer and Cross Entropy Loss.

  - **Training Loss:**  

    - Epoch 1: 0.16766  

    - Epoch 2: 0.07790  

    - Epoch 3: 0.05842  

    - Epoch 4: 0.04947  

    - Epoch 5: 0.04059  

    - Epoch 6: 0.03726  

    - Epoch 7: 0.03220  

    - Epoch 8: 0.02918  

    - Epoch 9: 0.02811  

    - Epoch 10: 0.02485  

    - Epoch 11: 0.02318  

    - Epoch 12: 0.02107  

    - Epoch 13: 0.01879  

    - Epoch 14: 0.01641  

    - Epoch 15: 0.01799

  - **Test Accuracy:** 99.24%

- **Model Saving & Download**  

  The trained model is saved as `CNN_Digit_Recognizer.pth`.  

  You can download the model from [this link](https://your-model-download-link.com).

## Requirements

- Python 3.x

- [PyTorch](https://pytorch.org/)

- [Torchvision](https://pytorch.org/vision/stable/index.html)

- [OpenCV](https://opencv.org/)

- [NumPy](https://numpy.org/)

- [Matplotlib](https://matplotlib.org/)

- Git (to clone the [HodaDatasetReader](https://github.com/amir-saniyan/HodaDatasetReader) repository)

## Setup and Usage

1. **Clone the Repository**  

   The notebook automatically clones the HodaDatasetReader repository using:

   ```bash

   !git clone https://github.com/amir-saniyan/HodaDatasetReader

   ```

   Make sure the dataset files (`Train 60000.cdb`, `Test 20000.cdb`, and `RemainingSamples.cdb`) are in the `DigitDB` folder.

2. **Run the Notebook**  

   - Open the notebook in [Google Colab](https://colab.research.google.com/).

   - Run all cells sequentially to load the data, visualize samples, train the CNN, and evaluate its performance.

   - The final cell saves the trained model.

3. **Download the Trained Model**  

   After training, download the saved model (`CNN_Digit_Recognizer.pth`) from [this link](https://your-model-download-link.com).

## Customization

- **Image Dimensions:**  

  Change `images_height` and `images_width` parameters to adjust the input image size.

- **Model Architecture:**  

  Modify the `CNN` class if you wish to experiment with different network configurations.

- **Training Parameters:**  

  Tweak the number of epochs, batch size, or learning rate to explore different training setups.

## License

This project is provided as-is without any warranty. You may modify and use it for educational or research purposes.