#Behavioral Cloning
This project contains the following files:
- drive.py: Python file for steering, uses model.h5
- model.h5: Trained model
- model.py: Pyhon file to train the model
The model structure is as follows:
Layer (type) |Output Shape
cropping2d_1 (Cropping2D) |(None, 66, 320, 3)
lambda_1 (Lambda) |(None, 32, 160, 3)
lambda_2 (Lambda) |(None, 32, 160, 3)
convolution2d_1 (Convolution2D) |(None, 8, 40, 16)
elu_1 (ELU) |(None, 8, 40, 16)
convolution2d_2 (Convolution2D) |(None, 4, 20, 32)
elu_2 (ELU) |(None, 4, 20, 32)
convolution2d_3 (Convolution2D) |(None, 2, 10, 64)
flatten_1 (Flatten) |(None, 1280)
elu_3 (ELU) |(None, 1280)
dense_1 (Dense) |(None, 512)
elu_4 (ELU) |(None, 512)
dense_2 (Dense) |(None, 1)
Total params: 723,569
While training the model, it came to my attention that the final model was better if dropout was skipped, so I opted for prunning the dropout layers (lines 205, 208)
Loss didn´t see to change after epoch 3-4, so I ended up using 10 total epochs of training.
There was no test split, only validation and training
Experimentaly, a smaller learning rate using an Adam optimizer worked better (line 212)
As per training data, the udacity dataset was used along with a personal dataset, obtained from 2 full laps around track 2, clockwise and counterclockwise.
Data aquisition was made with keyboard and a gamer friend :]
Data was augmented by using the 3 cameras, shifting the steering angle for right and left for 0.25. (Lines 78-94)
As for trying to alleviate the 0 steering bias, 90% of the 0 steering data was deleted after augmentation (Lines 53-76)
Image processing was done as follows:
- Image cropping (line 174)
- Image resizing (Lines 182-184)
- Image normalization (Line 198)
