Convert Trained YOLO Model to Tensorflow Lite

[andrew-aj]

What needs to change?

As part of our move to Sub9, we need to get our vision models compatible with the computer. Traditionally, running YOLO requires a GPU which we have on Sub8 to run our ML pipeline on. However, Sub9 is a smaller form factor and as a result we cannot fit a GPU. Running our vision models just on the CPU would be too intensive but we have a potential work around. We bought an M.2 TPU accelerator here. However, it only runs TensorFlow Lite whereas our current YOLO model uses PyTorch. There is a potential work around where we can convert PyTorch models to TensorFlow Lite. This task is to determine if this is possible. Here is a link that you can use to start from: converting models.

How would this task be tested?

Contacts

• [ ] We need help from the mechanical team.
• [ ] We need help from the electrical team.
• [ ] We need help from Dr. Schwartz or other faculty.
• [ ] We need help from a company or an organization.
• [ ] We need help from another UF staff member or organization (ex, facilities).