Machine Learning project to classify car models by photos. Inspired by fastai course & code. So far this is just the first version and a proof of concept. Current version classifies cars by 8 categories, which represents 8 generations of Volkswagen Golf model.

Try it! (on mobile or desktop)

All parts of this application are written in python in form of Jupyter Notebooks. Further I will describe the main steps of this project and point out which notebooks are responsible for them.

• Data
  • Gathering
  • Splitting
• Training
  • Model
  • Accuracy
  • Example of Predictions
• Application
  • Deployment
• Main Dependencies

Data for this project was collected from Internet. Partially using scripts and search engines (Bing & DuckDuckGo). But mostly by searching & downloading by hand.
1500 images there collected in total. Approximately 190 images per category (so far there are 8 VW Golf generations). In addition, 67 photos were collected on the street with the iPhone camera.

The training data is not a part of this repo.

The data was then split in 3 sets: training (67%), validation (19%) & test (14%).

A pre-trained ResNet50 model was selected for the training. Then it was fine tuned using training data for 100 epochs using MacBook Pro (2018, i7, 16 GB) and only CPU (the training took about 6 hours). A standard set of augmentations from fastai library were applied to the training data.

You can download the model here.

Validation set - 96%
Test set - 95%
Own iPhone photos taken on street - 88%

A web application is build using Gradio library.

• fastai
• fastbook
• Jupyter Notebooks
• Voilà
• Gradio
• macOS