The mnist-digit-classifier is a collection of neural network models—Linear, Convolutional, and ResNet—designed for the classic MNIST digit classification task using PyTorch. This project showcases data handling, model training, and evaluation processes.
Instead of utilizing the MNIST dataset from torchvision.datasets, this project employs .csv files containing digits in the format label, pixel_1, pixel_2, ... pixel_784 sourced from Joseph Redmon. These .csv files are then made into Datasets that are passed to the DataLoaders.
Additionally, the project includes a small custom dataset of 130 digits (inputs were made with a trackpad or a stylus) drawn by 4 different people in .png format, which I collected and annotated. I was curious to see how the models would perform on data with different characteristics (e.g. handwriting style, digital creation) compared to the original MNIST dataset. This custom dataset is normalized and formatted the same way as the original MNIST dataset to ensure uniformity.
-
Clone the repository:
git clone [email protected]:andrasmatyassy/mnist-digit-classifier.git cd mnist-digit-classifier
-
Create a virtual environment (optional but recommended):
python -m venv .venv source .venv/bin/activate # On Windows: .venv\Scripts\activate
-
Install dependencies:
pip install -r requirements.txt
-
Download the MNIST dataset CSV files from Joseph Redmon and place them in the
data/mnistdirectory.- Training set: https://pjreddie.com/media/files/mnist_train.csv
- Test set: https://pjreddie.com/media/files/mnist_test.csv
The project uses a config.yaml file to manage configurations. Modify this file to adjust paths, hyperparameters, and other settings.
To run the project, use the following command:
python main.pyIf mode.train is set to true in the config.yaml file, the models will be trained and the intermediate results will be printed to the console.
Final evaluation is done on the custom dataset of handwritten digits. The evaluation of the models is saved to logs/test.log.
While the models perform well on the original MNIST test dataset, the performance drops significantly when evaluated on the custom dataset:
| Model | Accuracy on MNIST | Accuracy on Custom Dataset |
|---|---|---|
| Linear | 0.98 | 0.63 |
| Convolutional | 0.99 | 0.72 |
| ResNet | 0.99 | 0.92 |
This project is licensed under the MIT License.