Behavioral Cloning Project - Udacity Self-Driving Car Engineer Nanodegree.
This project is a part of the Udacity's Self Driving Car Nanodegree. The main focus is to develop a deep neural network capable to drive a vehicle on a simulated track. The input data is generated from human behavior using the same simulator.
model.py - Main script containing neural network.
datasetProcessor.py - Dataset preparation and training data augmentation.
model.h5 - Trained model that is able to complete successfully a track. Please, download it from here (1.32 GB).
run1.mp4 - The video file containing 2 successfull laps completed in an automonuous driving mode. Please, download it from here (43.9 MB, 320 X 160).
The environment can be setup following this link: CarND Term1 Starter Kit
More detailed project description is available here.
Training data can be downloaded from here. It is important to remove a first descriptive line in a csv file.
The neural network architecture used in this project follows, in general terms, the Nvidia's DNN described in this blog post. A bit simplified convolutional neural network shown below was enough to get a track completed. Network's implementation using Keras is the following:
model = Sequential()
model.add(Lambda(lambda x: x/255.0 - 0.5, input_shape=(160, 320, 3)))
model.add(Cropping2D(cropping=((50, 20), (0, 0))))
model.add(Conv2D(24, (5, 5), strides=(2, 2), padding='same', activation='relu'))
model.add(Conv2D(48, (5, 5), strides=(2, 2), padding='same', activation='relu'))
model.add(Conv2D(64, (3, 3), strides=(1, 1), padding='same', activation='relu'))
model.add(Conv2D(64, (3, 3), strides=(1, 1), padding='same', activation='relu'))
model.add(Flatten())
model.add(Dense(1000, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(100, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(50, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(10, activation='relu'))
model.add(Dense(1))
The Adam optimizer with learning rate 0.0001 was used.
Adam = optimizers.Adam(lr=learning_rate)
model.compile(loss='mse', optimizer=Adam)
The code above was transformed to the following:
Layer (type) Output Shape Param #
=================================================================
lambda_1 (Lambda) (None, 160, 320, 3) 0
_________________________________________________________________
cropping2d_1 (Cropping2D) (None, 90, 320, 3) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 45, 160, 24) 1824
_________________________________________________________________
conv2d_2 (Conv2D) (None, 23, 80, 48) 28848
_________________________________________________________________
conv2d_3 (Conv2D) (None, 23, 80, 64) 27712
_________________________________________________________________
conv2d_4 (Conv2D) (None, 23, 80, 64) 36928
_________________________________________________________________
flatten_1 (Flatten) (None, 117760) 0
_________________________________________________________________
dense_1 (Dense) (None, 1000) 117761000
_________________________________________________________________
dropout_1 (Dropout) (None, 1000) 0
_________________________________________________________________
dense_2 (Dense) (None, 100) 100100
_________________________________________________________________
dropout_2 (Dropout) (None, 100) 0
_________________________________________________________________
dense_3 (Dense) (None, 50) 5050
_________________________________________________________________
dropout_3 (Dropout) (None, 50) 0
_________________________________________________________________
dense_4 (Dense) (None, 10) 510
_________________________________________________________________
dense_5 (Dense) (None, 1) 11
=================================================================
Total params: 117,961,983
Trainable params: 117,961,983
Non-trainable params: 0
_________________________________________________________________
At the training stage, the neural network above is feeded with a data by generator. The generator function selects randomly central, left, or right picture from one string of the training dataset. All training data augmentation is done by datasetProcessor.py.
The training dataset is composed by the following pictures:
- original pictures
- pictures flipped horizontally
- original pictures with random brightness
- flipped pictures with random brightness
Training was performed with the following parameters:
- epochs = 7
- learning_rate = 0.0001
- batch_size = 256
An example of training and validation performance of the model is depicted in a Figure below.
Files model.py, model.h5, datasetProcessor.py, and run1.mp4 are distributed under MIT license.
Please refer to Udacity regarding all other supporting materials.