On a previous post I write about a work project using Tensor Flow 2.0. Since then I completed an initial round of training of the Inception v4 model with 145,000 images. Over the course of the past week I identified and classified 64,000 real-world images taken from our fielded prototype this past summer to use for testing.
I was very pleased with a 95% accuracy on this test data. After some inspection I determined that many of the errors were caused by misclassifying test images. A clear case of human error, that once remedied brought the accuracy closer to 98%.
At this point I began loading an office prototype, RPI0W, with the Tensorflow 2.0 libraries and related application software. After 3 minutes of execution the RPI0W froze. At this point I am unsure if it is a TensorFlow 2.0 issue or an issue with computing resources unable to handle the much more complex machine learning models. This summer we were using TensorFlow 1.14 on the RPI0W with a model that had at most a dozen layers. The Inception v4 model has 150+ layers.
This week will entail some troubleshooting to determine where the problem lay. Hopefully I can determine how to execute this model on our prototype as it promises to be far more accurate than the one we deployed this past summer.
