Car Behavioral Cloning using Pytorch

Overview

I create this project after watching Siraj's video about how to simulate a self-driving car. Instead using Keras to build a model, I want to use Pytorch for the model and training.

The repo is inspired by naokishibuya. I adapt his augmentation methods.

Video demo:

Youtube link

Challenges:

• I think the most challenge is generating enough data for your model. For some tricky curves on the road, you need to creat more data.
• Building a model in Pytorch is not as straighforward as in Keras. You need to understand the framework and how it processes data first.
• Need to create a Dataloader for your own data.
• Re-use as much code as possible.

What has been done so far

• Created the Dataloader for car dataset
• Created the CarModel for training with the dataset. The model is based on The NVIDIA model.
• Created another Simple Car Model by reducing the number of Convolutional and Linear layers.
• Modified drive.py so that it can load the Pytorch model.
• After training Simple Model around 15 epochs, the model can drive a car for one full lap without any crashing.
• I save two model, the first model is saved when the validation loss is lowest, the second model is saved after training phase stop. As I notice, the second model is quite smoother than the first one.
• Working models are put inside working_models dir

Model summary:

  (module): CarSimpleModel (
    (conv_layers): Sequential (
      (0): Conv2d(3, 24, kernel_size=(3, 3), stride=(2, 2), bias=False)
      (1): ELU (alpha=1.0)
      (2): Conv2d(24, 48, kernel_size=(3, 3), stride=(2, 2), bias=False)
      (3): MaxPool2d (size=(4, 4), stride=(4, 4), dilation=(1, 1))
      (4): Dropout (p = 0.25)
    )
    (linear_layers): Sequential (
      (0): Linear (3648 -> 50)
      (1): ELU (alpha=1.0)
      (2): Linear (50 -> 10)
      (3): Linear (10 -> 1)
    )

Notes

• Training the neural network

• After trained, the network is able to generate steering commands from the video images of a single center camera.

• Testing the model with the Simulator

Dependencies

• You can install all dependencies by running conda:

conda env create -f environments.yml

• Note: I've removed tensorflow in this file
• After that, you need to install Pytorch
• To run up the server, you need to download Udacity Self-Driving Car Simulator. (using Autonomous Mode when trying your model)