This week’s main goal was to collect 100,000 audio samples with Bruce (our makeshift echo locator bat) and work on the neural network in hopes of getting some decisive results. Sadly, but humorously, we managed to run into every single possible problem before we even got to training and testing the neural net.
One of the problems we encountered was that the bat ears were producing weak audio signals that would fall below our alignment threshold. Our three possible solutions to this were:
- Rotate the mics or reposition them at an angle at which the sound is pointing towards to
- Take off the bat ears completely, since we established earlier that the sound is stronger without them
- And lastly, our ultimate solution, increase the gain on the microphones. Unfortunately, at this point we had to abandon some 20,000 samples we took previously and start collecting once more with the new settings.
A second problem we had was opening and processing the samples we took previously, getting errors with our h5 files such as “Unable to open file: Can’t open ‘audio’” and “Unable to truncate file”. Apparently, after spending a whole day learning about h5 files, the process was crashing once it was trying to take out the audio and depth from the files. Some of the ways we approached this problem were:
- We went through the h5 code left to us by Dr. Sprague and tried to see if there is a fault with it in saving our files as .h5 files, but we could find nothing wrong with it
- Try taking small samples of data and process it as an npz file and it worked, so that led us to try and:
- Change the h5 data collecting code to save as npz instead. After many hours of work, we realized that it would be impossible since we could not understand or visualize the resizing of the matrixes going on with the h5 files, thus we could not recreate it with an npz version
- After a whole day of obsessing over h5 files, our solution came from where we did not expect. Because going forward with any type of changes did not work, we decided to take steps backwards and undid every change. And ~Eureka~ it worked. We have no clue why it did not work originally, resulting in all types of errors, but now it seemed to be pleased with us and let us move on from this and onto greater things.
A third problem we encountered was collecting data and getting remote access to it through Zemenar. We collected and then stored our data on a flash drive instead of saving in to Zemenar because it took less time (saving it Zemenar took about 2 hours per 3-4,000 samples). From here we had problems getting access to it, and so some possible solutions we found were:
- Attempt to run a neural network from a Linux lab machine, which did not work because the GPU ran out of memory
- Let data copy over to Zemenar overnight
- Process the data on a computer that has access to our flash drive and then copy the resulting file to Zemenar (takes less than a minute) and then run the neural network from there. This seemed to solve our problem.
After finding solution to all those problems and managing to take good sample data from the lab room we were currently in, we decided to move on to another classroom to record more. But because our day of misfortune was not over just yet, the data Bruce collected was just a big patch of color. After giving it a little break and rechecking all settings and connections, Bruce came back to normal. I have an inference that the problem occurred due to some cable getting dislodged a little due to our movements, so maybe taping them would be good. Our laptop was giving us a hard time as well, freezing multiple times.
As the week draws to an end, we are hoping most of our problems are past us and we can focus on finishing up collecting the 100,000 samples, collect samples with a specific object from different visual perspectives, and use it to get some conclusive results from out neural networks.