Since we ran out of AWS credits, I turned off the EC2 instance last week. However, this week we received more so I started a new EC2 instance. It turns out that when you start an AWS instance, it keeps the settings of the person who last used it. My AWS instance had a lot of problems because it wasn’t updated to the latest EC2 version, and had libraries which were different versions from the ones that I used, so my old code couldn’t run on the EC2. After reupdating the entire EC2, and uninstalling and reinstalling various versions of the libraries I was using, I was able to get my network to train, but I could not download the model. Instead I had to save the model as a checkpoint file, and then convert the checkpoint file to a saved_model object on my own computer.
I retrained the ML model with improvements to the convolutional neural network and the larger data set. I added several convolution layers and a batch normalization layer, as well as 600 more images (times 2 after flipping them). I ended up getting 84% accuracy on the validation set.
I also helped Philip get libraries set up on the Jetson and we were able to successfully run our computer vision + machine learning code on it.
Unfortunately, I fell sick during the latter half of this week.