Testing

Coral Accelerator

Google provides a number of examples to test the Coral accelerator boards. The example we chose to run which is most similar to the work that was being done in PyTorch is the classification example, which classifies an image of a bird to one of a few species. We found that after the initial classification (loading the model into the Coral's memory),  it took single digit milliseconds to complete. 

Image of test being run to display model times and speedup

Switching to TensorFlow

Based on the data we collected, and the testimony of members who worked on the project in previous semesters, we believe that switching to TensorFlow to be able to take advantage of the accelerator board will provide a significant improvement in model performance compared to running the model on the Raspberry Pi's CPU only. We are still figuring out if it's possible to convert the previous YOLOv7 based model to TensorFlow, or if we will need to re-train from scratch.

Using Docker Containers

In order to streamline testing and development we have been utilizing Docker containers in order to make it so that any user accessing the container will have access to the preloaded imports and other programs downloaded into the container. Doing so has made it so we are able to easily work on our code despite working on different devices and operating systems.