Lake Track | Jungle Track |
---|---|
YouTube Link | YouTube Link |
In this project, I use a neural network to clone car driving behavior. It is a supervised regression problem between the car steering angles and the road images in front of a car.
Those images were taken from three different camera angles (from the center, the left and the right of the car).
The network is based on The NVIDIA model, which has been proven to work in this problem domain.
As image processing is involved, the model is using convolutional layers for automated feature engineering.
model.py
The script used to create and train the model.drive.py
The script to drive the car. You can feel free to resubmit the originaldrive.py
or make modifications and submit your modified version.utils.py
The script to provide useful functionalities (i.e. image preprocessing and augumentation)model.h5
The model weights.environments.yml
conda environment (Use TensorFlow without GPU)environments-gpu.yml
conda environment (Use TensorFlow with GPU)
Note: drive.py is originally from the Udacity Behavioral Cloning project GitHub but it has been modified to control the throttle.
You need an anaconda or miniconda to use the environment setting.
# Use TensorFlow without GPU
conda env create -f environment.yml
# Use TensorFlow with GPU
conda env create -f environment-gpu.yml
Or you can manually install the required libraries (see the contents of the environment*.yml files) using pip.
Start up the Udacity self-driving simulator, choose a scene and press the Autonomous Mode button. Then, run the model as follows:
python drive.py model.h5
You'll need the data folder which contains the training images.
python model.py
This will generate a file model-<epoch>.h5
whenever the performance in the epoch is better than the previous best. For example, the first epoch will generate a file called model-000.h5
.
The design of the network is based on the NVIDIA model, which has been used by NVIDIA for the end-to-end self driving test. As such, it is well suited for the project.
It is a deep convolution network which works well with supervised image classification / regression problems. As the NVIDIA model is well documented, I was able to focus how to adjust the training images to produce the best result with some adjustments to the model to avoid overfitting and adding non-linearity to improve the prediction.
I've added the following adjustments to the model.
- I used Lambda layer to normalized input images to avoid saturation and make gradients work better.
- I've added an additional dropout layer to avoid overfitting after the convolution layers.
- I've also included ELU for activation function for every layer except for the output layer to introduce non-linearity.
In the end, the model looks like as follows:
- Image normalization
- Convolution: 5x5, filter: 24, strides: 2x2, activation: ELU
- Convolution: 5x5, filter: 36, strides: 2x2, activation: ELU
- Convolution: 5x5, filter: 48, strides: 2x2, activation: ELU
- Convolution: 3x3, filter: 64, strides: 1x1, activation: ELU
- Convolution: 3x3, filter: 64, strides: 1x1, activation: ELU
- Drop out (0.5)
- Fully connected: neurons: 100, activation: ELU
- Fully connected: neurons: 50, activation: ELU
- Fully connected: neurons: 10, activation: ELU
- Fully connected: neurons: 1 (output)
As per the NVIDIA model, the convolution layers are meant to handle feature engineering and the fully connected layer for predicting the steering angle. However, as stated in the NVIDIA document, it is not clear where to draw such a clear distinction. Overall, the model is very functional to clone the given steering behavior.
The below is a model structure output from the Keras which gives more details on the shapes and the number of parameters.
Layer (type) | Output Shape | Params | Connected to |
---|---|---|---|
lambda_1 (Lambda) | (None, 66, 200, 3) | 0 | lambda_input_1 |
convolution2d_1 (Convolution2D) | (None, 31, 98, 24) | 1824 | lambda_1 |
convolution2d_2 (Convolution2D) | (None, 14, 47, 36) | 21636 | convolution2d_1 |
convolution2d_3 (Convolution2D) | (None, 5, 22, 48) | 43248 | convolution2d_2 |
convolution2d_4 (Convolution2D) | (None, 3, 20, 64) | 27712 | convolution2d_3 |
convolution2d_5 (Convolution2D) | (None, 1, 18, 64) | 36928 | convolution2d_4 |
dropout_1 (Dropout) | (None, 1, 18, 64) | 0 | convolution2d_5 |
flatten_1 (Flatten) | (None, 1152) | 0 | dropout_1 |
dense_1 (Dense) | (None, 100) | 115300 | flatten_1 |
dense_2 (Dense) | (None, 50) | 5050 | dense_1 |
dense_3 (Dense) | (None, 10) | 510 | dense_2 |
dense_4 (Dense) | (None, 1) | 11 | dense_3 |
Total params | 252219 |
- the images are cropped so that the model won’t be trained with the sky and the car front parts
- the images are resized to 66x200 (3 YUV channels) as per NVIDIA model
- the images are normalized (image data divided by 127.5 and subtracted 1.0). As stated in the Model Architecture section, this is to avoid saturation and make gradients work better)
For training, I used the following augumentation technique along with Python generator to generate unlimited number of images:
- Randomly choose right, left or center images.
- For left image, steering angle is adjusted by +0.2
- For right image, steering angle is adjusted by -0.2
- Randomly flip image left/right
- Randomly translate image horizontally with steering angle adjustment (0.002 per pixel shift)
- Randomly translate image vertically
- Randomly added shadows
- Randomly altering image brightness (lighter or darker)
Using the left/right images is useful to train the recovery driving scenario. The horizontal translation is useful for difficult curve handling (i.e. the one after the bridge).
The following is the example transformations:
Center Image
Left Image
Right Image
Flipped Image
Translated Image
I splitted the images into train and validation set in order to measure the performance at every epoch. Testing was done using the simulator.
As for training,
- I used mean squared error for the loss function to measure how close the model predicts to the given steering angle for each image.
- I used Adam optimizer for optimization with learning rate of 1.0e-4 which is smaller than the default of 1.0e-3. The default value was too big and made the validation loss stop improving too soon.
- I used ModelCheckpoint from Keras to save the model only if the validation loss is improved which is checked for every epoch.
As there can be unlimited number of images augmented, I set the samples per epoch to 20,000. I tried from 1 to 200 epochs but I found 5-10 epochs is good enough to produce a well trained model for the lake side track. The batch size of 40 was chosen as that is the maximum size which does not cause out of memory error on my Mac with NVIDIA GeForce GT 650M 1024 MB.
This track was later released in the new simulator by Udacity and replaced the old mountain track. It's much more difficuilt than the lake side track and the old mountain track.
I used the simulator to generate training data by doing 3 to 4 rounds. Also, added several recovery scenarios to handle tricky curves and slopes.
I felt that the validation loss is not a great indication of how well it drives. So, I tried the last several models to see which one drives the best. For this, I set the save_best_only to False (use -o false
for model.py), and I used 50 epcohs (Use -n 50
).
The model can drive the course without bumping into the side ways.
- NVIDIA model: https://devblogs.nvidia.com/parallelforall/deep-learning-self-driving-cars/
- Udacity Self-Driving Car Simulator: https://github.com/udacity/self-driving-car-sim