Status Report #10
Joel :
Finally putting the neural network on the Pi and completely integrating to see what would happen was an entire process. In fact, just to install tensorflow on the raspberry Pi was a huge problem. Originally we didnt have enough space on our sd card and as a result, we had to first buy a bigger sd card. We found out that some SD cards also allow for faster read and write access. So we decided to go large for both the faster read/write access and for the extra space to download all the components that we needed. The Neural Network was trained on my laptop, and the model weights were saved to a file and then sent to the pi where it would be loaded up when we ran our main function. Unfortunately the first time we ran it after finally successfully installing tensorflow. The first time we ran everything, our PI completely bricked. It was strange and something we didn’t account for. The PI wouldn’t even turn on after rebooting, restarting, and reflashing. So this required us to order new Pis from amazon with 1 day shipping in order to have a few backups prepared. We also needed a new plan for integration in order for this to work smoothly. We came up with 2 ideas. First, create a smaller but still accurate enough which wouldn’t take up as much space. The second thing we did was also install something called tensorflow lite on the raspberry pi, which made the computation time for the model much faster. The combination of both these things really helped and made our integration smoother.
Chinedu:
This week I focused on finalizing the circuit and documenting the final setup of the G-Lock circuit. We had to make the final design decision on how to power the system and we decided that we wanted to use wall power to power the circuit and the RasPi. The fact that the door was static meant that there was no real need for portable power. In addition, I was able to finalize the LED circuitry and the strike circuitry. In this week I also had to prepare our team’s final presentation slides. In these slides, I detailed the solution approach which includes our components and the design of our neural network. In this presentation, I focused on the metrics and validation of our G-Lock and reporting if our team met our design requirements. Giving this presentation was a great learning experience in delivering data-driven presentations. I am happy our group met most of the success criterion.
Omar:
Coming up to the final week, we had to make some pretty big decisions with our model and pipeline. Everything was coming together, but we still needed to make the final connections between our different subsystems and make sure everything was working. A lot of this week was spent assembling components and putting them onto the door with the correct wiring. We also got the connection fully established between the new neural network and the rest of the system, so we were ready to test once the electric strike/raspberry pi circuit parts were integrated. When everything was finally assembled, and we were ready to go, we started testing. Our test efficiency was shaky in the beginning. Chinedu was being recognized perfectly every time, but I couldn’t even get recognized once by the door. Both parts of the neural network algorithm recognized me, but it just wasn’t certain enough. We played around with these values and made adjustments to our neural network weights.