From a simulator provided by Udacity, ones should generate training data to train a neural network to predict steering angles. Simulator's output returns image views from center, left and right of the car. In addition, an csv file containing:
- Steering Angles
- Break
- Throttle
- Speed
The implemented model was inspired on NVIDIA's architecture as presented on this paper:
Although the model above has provided important details regarding its definition, it's not mentioned important information like hyper-parameters and non-linearity functions.
In order to achieve a fast implementation of the model and quickly tweak architecture, Keras has been used. Keras allows you to modify a model by just modifying A few lines of code.
The number of training data is to high to be fully loaded in to memory, which demands images being loaded into batches to be normalized before the insertion into training pipeline. Preprocessing is carried out by generator() function, yielding data to training.
No data augmentation was performed for this project.
- Activation: RELU's after every layer
- Pooling Layers: After Every Convolutional Layer
- Strides = 2x2
Optimizer: Adam optimizer - http://arxiv.org/abs/1412.6980v8 Metric: As the problem is characterized as a regressor, the correct metric is "mean squared error" which computes the mean of the squered difference between the predicted and correct value:
Hyper Parameters:
- Learning Rate: 0.0001
- Batch Size: 128
- Number of Epochs: 2
- Dropout: 0.4
model_dict = {'in_shape': (66, 200, 3), # For tensorflow backend use shape as (row, cols, chan) 'nb_filter_1': 24, 'nb_row_1': 5, 'nb_col_1': 5, 'stride_1': (2, 2), 'z_1': 'relu', 'pad_1': 'valid', 'nb_filter_2': 36, 'nb_row_2': 5, 'nb_col_2': 5, 'stride_2': (2, 2), 'z_2': 'relu', 'pad_2': 'valid', 'nb_filter_3': 48, 'nb_row_3': 5, 'nb_col_3': 5, 'stride_3': (2, 2), 'z_3': 'relu', 'pad_3': 'valid', 'nb_filter_4': 64, 'nb_row_4': 2, 'nb_col_4': 2, 'stride_4': (1, 1), 'z_4': 'relu', 'pad_4': 'valid', 'nb_filter_5': 64, 'nb_row_5': 2, 'nb_col_5': 2, 'stride_5': (1, 1), 'z_5': 'relu', 'pad_5': 'valid', 'drop_p': 0.4}
The inital dataset consisted of aproximatelly ten thousand samples, which was assessed as not sufficient for providing enough information to successfully train the neural net. To achieve a sufficient number of images, several sessions of running through cenarios one and two had been done, most importantly recording the session at specific points of the track to scape from collision path.
Data Set statistics: Data set has been splitted into 90% for training data and 10% for testing and validation:
- Train Dataset: 224.969
- Validation: 18747
- Test Size: 6250
- Training time: 15 minutes with Geforce 1060 6GB GPU