ml-pipeline

🔑 Key Takeaways:

• Built a Convolutional Neural Network (CNN) to train a computer vision model that recognizes digits drawn by the user.

• Used PyTorch's transforms v2 to preprocess the dataset — including converting images to tensors, scaling pixel values, and normalizing them — to improve training performance.

• Trained the model on the MNIST dataset of handwritten digits, using the training set and validating it against the test set.

• After 15 epochs, the model reached approximately 95% accuracy.

• Created a Streamlit web app with a drawable canvas that allows users to sketch a digit and get a prediction from the trained model in real-time.
  The use of proper transforms during inference significantly improved prediction quality.

• Next step: I’m working on deploying the model using FastAPI for the backend and integrating a PostgreSQL database to log user predictions and track analytics.
  I'm also exploring ways to further improve the model's performance and generalization.