Project 3: Behavioral Cloning

The aim of this project was to train a deep learning model that can teach a car of how to drive around a track in a udacity developed simulator. The data used to train this model is generated using the udacity simulator.

Files

model.py: Python script that imports data, preprocesses data, trains model and saves model
model.json: Contains Model Architecture
model.h5: contains model weights
drive.py: A modified python scipt that is used to drive in udacity simulator using the trained model and a given dataset
model_experiments.ipynb: Overview of various experiemnts tried to get a working pipeline
To run in autonomous mode use the below command:

python drive.py model.json

DataSet

I collected two datasets using udacity simulator. One data set was collected using mouse and other was collected using arrows
I also utilized dataset provided by udacity
I also learned through various experiemnts that only using data that has straight driving and waring off the road was not sufficient to get a successful training model. The training data also need flavour of recovery samples whereby the car recovers when it goes off the track ( crosses the yellow border or red markers on turns). Thanks to Pkern, I utilized his recovery data set only along with the above mentioned dataset.

Method utilized for final pipeline

Preprocessing data

Data engineering was by far the most important part of this assignment.
I started with using the cropped images as advised in the class but that didn't get me to a working model
I utilized udacity dataset which was displined driving in the center of the lane. This dataset was far away from real word scanerio whereby there might be times when car crosses the yellow shoulder line. Example of center lane driving shown below:
I introduced flipped images which immitated right turns on the circuit. This was required as most of the data collected represented left turns. This enabled us to get more balanced dataset. Example image flipping shown below:

I also added more samples of left and right turns so that I can get a balanced dataset which immitates left, right and straight driving.
I utilized recovery data recorded by pkern which moves the car from left to center of lane and right to center of lane. Example shown below:

Thanks to blog post written by Vivek Yadav, Denise James, https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html, it provided sufficient guidance to move further whenever I was stuck with a problem

Model Architecture

I tried different model Architecture
A subset of Nvidia Behavior cloning model used by them in the paper
A modified version of vgg_net()
A Convolutional Neural Network (model_cnn) model that I came up with

All this models ran successfully around the the circuit thanks to the enormous amount of time spent on data engineering. I finally ended up using the CNN model that I came up with as it had less parameters to tune.

Below is the model used:

Hyper parameters

Below are the hyper parameters used for training the model

l2_reg = 0.001
keep_prob = 0.5
batch_size = 64
min_delta=1e-4
patience=4
nb_epoch = 100
beta2 = 0.999
beta1 = 0.9
epsilon = 1e-08
decay = 0.01
lr = 0.001
I also used keras's call back functions so that the training can be killed is not further improvements is added. To do so a patience factor of 4 with a minimum improvement parameter of 0.0001.

Results

The video of the results can be found

The results attained was tested on a different dataset that I collected which was not used in training or validation.

patelmiteshn/BehavioralCloning