Week 19. Testing and exploring the pilot code.

1 minute read

To Do

  • New estructure in Neural Behaviors repo.
  • Execute a neural pilot and pilot solution.

Progress

1. New estructure in Neural Behaviors repo

In the process of restructuring the code mentioned in the previous weeks these changes are being taken to the Neural Behaviors repository.

The repository has been subdivided (for the time being) into three lines of research. One, the one already known and in which we are working the last weeks and that part of the final work of Vanessa Fernandez called: “Vision-based end-to-end using Deep Learning”.

In this repository you can see the structure that is being created to include in it the solutions of this end of master’s work as well as others that may be included.

2. Execute a neural pilot and pilot solution

Three solutions to the restructured code have been presented here:

  • Manual pilot. Solution taken from the code of the master in artificial vision.

  • Pilot classifier. Using deep learning, values of ‘v’ and ‘w’ distributed in the number of classes corresponding to the loaded model (5 or 7) are returned to the network output.

  • Pilot Return: Using also neural networks, it is obtained to the output numerical values that indicate the amount of v or w that must be applied to the car to follow the line.

You can see the solutions in the following gallery:

Manual Pilot Classificator Pilot Regressor Pilot
Different solutions to the follow-the-line exercise. With and without neural pilot.

It can be seen as a result that the solution is very similar. For more information visit the Neural Behaviors repository.

Working

In order to gain confidence and security with neural network training we will train a network from the beginning with an existing dataset. As a result we expect to obtain a solution just like the existing one.

Learned

Seeing the different configurations and possibilities offered by the repository we will focus the project to create a tool that allows switching between different networks and offers quality metrics to compare results. It can be said that what we have learned is that, from something apparently closed as it can be to solve the problem for the final master’s work, we can build a useful tool for future users, students, companies, etc.

You May Also Enjoy