https://review.udacity.com/#!/rubrics/432/view
Training the model was developed with the following steps:
- Load the data from the log file, which has 8036 rows with 'center', 'left' and 'right' image paths and the floating value for the steering in degrees
- Shuffle and split the log file
- Load and preprocess each batch of data using a generator
- Each training batch uses a generator for the validation set
- The validation set does not flip the images.
The data contains a log file and images for the left, center and right cameras, including the steering value.
We can see that the data has a bias towards the right by using a histogram.
To train the network, we use three laps with the data provided with the project.
http://alexlenail.me/NN-SVG/LeNet.html
We develop three models for the network:
- A simple network to test out the different configurations of the Keras layers
- LeNet5 configuration to train and validate. We use it to learn how the different configurations can affect the results
- The Nvidia model adding dropout layers to lower the overfitting
- Finally, we use the test data to validate the loss of the model
To reduce the overfitting of the model we use the following strategies:
- Since the steering data has a bias towards the right we flip the images from right to left and get the inverse
steering measure by multiply by -1
- We randomly set the gamma of the images
- Use the left and right cameras and compute a +/- 0.25 correction for the steering to the right and left cameras, so for each frame we have three images and three different steering measures. This helps the network to know how to recover when driving to the sides
- Use dropout layers of 0.25 to the model, improving the loss measure, and reducing overfitting.
- We use batch_size = 32 x 2(side cameras)+1(flip image)
- We use Adam optimizer with a learning_rate = 0.0003 that gives a proper loss improvement per epoch
- For the loss, we use Mean Absolute Error since we are solving a regression problem
- To learn what parameters fit best we use 15 epochs. However, most of the time 10 was the best number, before the loss begins to increase again
- We use a dropout 0.25 to improve the loss measure
The data provided was divided into 60% training, 20% validation, and 20% test data.
To improve performance and readability we use the text data only for shuffling and splitting, afterwords the generators only have to read the images from the paths specified.
The data was load into memory using generators to load only the batch images into memory.
The Nvidia model architecture has the following design. We added dropout layers and obtain an architecture as described here:
Total params: 348,219
Trainable params: 348,219
| Layer | Parameters | Activation | Output Shape | Weight parameters |
|---|---|---|---|---|
| Lambda | Normalization | (160, 320, 3) | 0 | |
| Cropping | 70,25 pixels | (65, 320, 3) | 0 | |
| Convolution (filters, kernel_size, stride) | 24,5,2 | Relu | (31, 158, 24) | 1824 |
| Convolution (filters, kernel_size, stride) | 36,5,2 | Relu | (14, 77, 36) | 21636 |
| Convolution (filters, kernel_size, stride) | 48,5,2 | Relu | (5, 37, 48) | 43248 |
| Dropout | 0.25 | (5, 37, 48) | 0 | |
| Convolution (filters, kernel_size, stride) | 64,3,1 | Relu | (3, 35, 64) | 27712 |
| Convolution (filters, kernel_size, stride) | 64,3,1 | (1, 33, 64) | 36928 | |
| Flatten | (2112) | 0 | ||
| Dense | 100 | (100) | 211300 | |
| Dropout | 0.25 | (100) | 0 | |
| Dense | 50 | (50) | 5050 | |
| Dense | 10 | (10) | 510 | |
| Dense | 1 | (1) | 11 |
The process of getting the data, preprocessed, loading, and feeding it to the network benefits from having good software engineering practices. However, the design of the network requires good science practices. Each training cycle is long from hypothesis to test it, improve it, and need to be good enough results.
Having said this, it is highly satisfactory to be able to develop the full cycle and see the results at the end of the process.
Having the car drive around the track is just part of a very complex task. Even so, we have not addressed the question of identifying other items such as automobiles and people to be able to avoid them.