This week I tested the solenoid and ordered a 12V DC Adapter and breadboard to help power the solenoid. I built the same circuit as the solenoid circuit diagram I drew last time.
I also scraped a few hundred more images off of Google to increase the size of our data set for Machine Learning (300 more images of raccoons, and 200 more images of lower body, which took like 3 hours to get). I’ll continue to find more images so that we can have at least 600 or 700 of each. I am also considering taking out the “squirrel” prediction, because the chance that we see squirrels is low and isn’t imperative towards our primary goal, which is to differentiate between cats and raccoons and humans.
Unfortunately, the $100 of AWS credits we had ran out, so I requested more AWS credits (and hopefully will receive them). We initially requested $150 of AWS credits several weeks ago, but one of the codes didn’t work so we only had $100 to work with. I won’t train anymore on the EC2 for now until we get more credits to work with. However, as soon we as have more AWS credits (or in the case that we don’t get any, we’ll use our budget),
Finally, I tested the Computer Vision + ML inference on the cat that we had at home, as well as on my own self. I’ve recognized a few patterns in the algorithm. If there are any long, leg-like objects of a solid color in the image, it will recognize that as the lower body of a human (for the most part). If there are any gray animals or if there is too much gray color in the image, it detects that as a raccoon, since the raccoon images I have are mostly gray. As of now, the predictions mostly depend on colors in the image, as I stated above. Additionally, the ML inference will return classification probabilities with 97%+ accuracy. In other words, the inference gives us extreme results, whereas having uncertain results (such as a classification probability of 60%) is actually more helpful. After researching online, having a larger data set will help alleviate this (as a larger data set will help alleviate most problems), as well as implementing k-fold (usually 10-fold) cross validation. What cross validation is, is shuffling the data set and retraining on it using random partitions of the data set as the training set and validation set. This is what I plan on doing by the next time I train (which should be by Wednesday April 10).
Next week I will focus on implementing k-fold cross validation, finding more images, and getting the CV and ML code to run using TensorRT.