/porsche_hackaton

2023 Porsche Engineering Prague Hackaton e2e training code

Primary LanguageJupyter Notebook

E2E learning for Porsche Hackaton

Concept

Collect a dataset of (image, steering_angle) pairs and train a conv-net to predict the steering angle from the image.

Data format

  • data/ folder, with .jpg as the image and .txt for the steering angle.
  • Tested on Sample dataset:
    • just unzip it in the project root and use the 'load_from_kaggle' dataset function (default)
    • Related blogpost

TODO

  • dataset class
  • inference code
  • training loop
  • visualization:
    • plot_sample function
    • plotting based on error
  • data augmentation to incentivize recovery behavior
  • dataset resampling to under-sample straights / over-sample turns
  • multi-image input to learn temporal context