Skip to content

PROJECT · ML / COMPUTER VISION

Tesla CNN — Keras/TensorFlow

A from-scratch convolutional neural network trained on a custom Tesla image dataset, using data augmentation to squeeze maximum signal out of a relatively small collection of images.

Python Keras TensorFlow Image augmentation

Problem & approach

The goal was to build a classifier focused on Tesla imagery, starting from a very small dataset. Rather than using an out-of-the-box architecture, this project defines a CNN from scratch so that the architecture, regularization, and augmentation strategy are all explicit design decisions.

Architecture & training

  • Stacked convolutional blocks with batch normalization and max pooling.
  • Dropout and L2 regularization to reduce overfitting on a small dataset.
  • Image augmentation (rotation, zoom, horizontal flips) applied via Keras generators.
  • Model trained and evaluated with accuracy and loss curves monitored over time.

What this project explores

This project is less about beating a leaderboard and more about understanding how architecture, augmentation, and regularization interact when you don’t have millions of labeled examples. It’s an experiment in shaping the model by hand and watching how each choice affects generalization.