Bengali.AI Handwritten Grapheme Classification Bronze Medal Solution

About The Project

This is my solution to the Bengali.AI Handwritten Grapheme Classification competition. I used a DenseNet-121 that receives a one-channel image and outputs 3 separate classifiers for the grapheme root, vowel diacritics, and consonant diacritics, respectively. The MixUp augmentation is used to improve the performance.

Result: Macro-average recall score of 0.9319 in the private leaderboard. Ranked 182 out of 2059 teams (bronze medal region).

Getting Started

To get a local copy up and running, follow these simple example steps.

Prerequisites

PyTorch (version 1.3.0)

Install using Anaconda:
```
conda install pytorch=1.3.0 -c pytorch
```

Installation

Clone the repo

git clone https://github.com/stephenllh/bengali_character.git

Change directory
```
cd bengali_character
```
Install packages
```
pip install requirements.txt
```

Usage

Change directory
```
cd bengali_character
```
Create a directory called input
```
mkdir input
cd input
```
Download the dataset into the folder
- Option 1: Use Kaggle API
  - pip install kaggle
  - kaggle competitions download -c bengaliai-cv19
- Option 2: Download the dataset from the competition website.
Run the training script
```
cd ..
python train.py
```
(Optional) Run the inference script
```
python inference.py
```

License

Distributed under the MIT License. See LICENSE for more information.

Contact

Stephen Lau - Email - Twitter - Kaggle

Project Link: https://github.com/stephenllh/bengali_handwriting

stephenllh/bengali-handwriting