Handwritten Bangla Character Recognition using CNN in PyTorch

About the dataset

The Original Dataset contains Handwritten Bangla Characters and Digits, a total of 84 classes where,

• 11 are Vowels
• 37 are consonants
• Rest are some of the Complex characters (consisting of 2 or more graphemes)

Instructions for Training the model

• Open CMD/Terminal and clone the repository using the command : git clone git@github.com:ayan-cs/handwritten-character-recognition-banglalekha
• Download the BanglaLekha Numerals dataset from the given link above and extract inside the repository folder. It is recommended not to make any change to the dataset folder.
• Preprocess the data by executing the Data_Preparation.ipynb notebook. This should create Train and Validation splits inside the parent dataset folder. (If you want to reverse the split and re-split the dataset again, a code snippet is available inside the notebook)
• Configure train_config.yaml file.
• Run the script on CMD/Terminal : python main.py train
• The trained model will be available inside Checkpoints folder and the plots will be saved inside Plots & Outputs folder.

Instructions for Inference/Prediction

• Open CMD/Terminal and clone the repository using the command : git clone git@github.com:ayan-cs/handwritten-character-recognition-banglalekha
• Make sure the data is preprocessed.
• Configure inference_config.yaml file. For demo, one trained ResNet-34 model has been provided.
• Open CMD/Terminal, run the command : python main.py inference
• The outputs will be available inside the Plots & Outputs folder.