Logo Lanfrica

wubeshetA/HahuNet

Domain:

natural language processing

Record type:

model
Creator:
wub
Host:
HahuNet is Convolutional Neural Network (CNN) to identify handwritten Amharic characters. # HahuNet HahuNet is a Convolutional Neural Network (CNN) to recognize handwritten Amaric characters. The model is trained on a dataset of 28x28 grayscale images with over 32,000 data points. HahuNet is a project that I started to learn about deep learning and computer vision. The name "Hahu-Net" is found by combining Ha and hu, which are the sounds of the first and second characters in the amharic alphabet and Net is a common suffix for neural networks. In Ethiopia Hahu (ሀሁ) is a common traditional name to refer to the entire Amharic alphabet. - Here is a picture of the Amharic alphabet: ## Model Architecture The model architecture is defined in the `HahuNet2` class. It consists of the following layers: - Convolutional layer with 16 filters of size 3x3, ReLU activation - Max pooling layer with pool size 2x2 - Convolutional layer with 32 filters of size 7x7, ReLU activation - Max pooling layer with pool size 2x2 - Convolutional layer with 64 filters of size 9x9, ReLU activation - Max pooling layer with pool size 2x2 - Flatten layer - Dense layer with 512 units, ReLU activation - Dropout layer with rate 0.5 - Output Dense layer with softmax activation The model uses categorical cross-entropy as the loss function and accuracy as the metric. ## Installation ```bash > git clone github.com > cd HahuNet/src > pip install -r requirements.txt ``` Run the flask app with the following command: ```bash > python app.py ``` ## Evaluation Currently the Model achieved 86.9% accuracy on the validation and test set. It also achieved 97% accuracy on the training set. I look forward to improving the model's performance on the validation and test sets. ## Future Work - Future work includes: - Improving the model's performance on the validation and test sets. - Build different digital tools that can be purposeful in the case of of identifying Amaric letters. This might include but not limited …