I'm Andras, a software engineer. I like automating stuff and tinkering with neural networks.
This project showcases the use of neural networks for classifying MNIST digits with PyTorch. It features Linear, Convolutional, and ResNet models, highlighting key aspects of data preprocessing, model training, and evaluation.
Instead of the usual torchvision.datasets, the project uses .csv files for pixel data, offering a unique approach to data handling. Additionally, a custom dataset of 130 handwritten digits in .png format is included to test how well the models handle images from a different data pipeline.
Key highlights:
- Data Handling: Transforms
.csvpixel data into PyTorch Datasets and DataLoaders. - Model Training: Easily configurable through a
config.yamlfile for hyperparameters and file paths. - Evaluation: Logs performance metrics, showing how model accuracy varies between the standard MNIST dataset and the custom dataset.
Built a config manager using the singleton pattern, complete with unit tests for robust model management.
Added environment variable to configure logging level.
