jjdblast / nvidia-autopilot-keras

Keras Implementation of End to End Learning for Self-Driving Cars by (Baris Kayalibay, Grady Jensen, Patrick van der Smagt)

End to End Learning for Self-Driving Cars

• Still work in progress, but basic model trained on GTA V dataset can be found:

  To Do:

  • Get proper driving data.
  • Inlcude the autonomy score.
  • Include model internal state viz.

About the Model

• Originally the model consists of:

The network consists of 9 layers, including a normalization layer, 5 convolutional layers and 3 fully connected layers. The input image is split into YUV planes and passed to the network.

• model.py contains the ConvNet, running the file will output a model to be used by train.py

Dependecy:

• Keras '1.2.0'
• Theano '0.9.0dev4.dev-RELEASE'

Dataset:

• The GTA V Dataset was used, previously it use to be hosted here.

• The dataset is messy often, but good enough for training/testing a model.

• If you have the dataset, load_deepdrive_data.py can be used to reduce the data into 128 and do any normalization desired for the model.

• trained model / weights, can be downloaded here

Training results for steering with Autopilot ConvNet:

• GPU (Trained 1080).

• Green is the original steering, and red is the Model's predictions.

the data contains weird incidents where the car is not going anywhere and suddent appearnce of other cars