Freespace segmentation using CUDA 11.1

I was able to follow the instructions to get CUDA 11.1 working with an Ampere GPU.
In testing out the Freescpace training app, I found that after moving to Tensorflow 2, some changes were needed in the training app.

I first had to swap out keras with tensorflow.keras:
from tensorflow.keras import backend as K

Then I had to use the compat.v1 for tensorflow
import tensorflow.compat.v1 as tf

I was able to get training up and running, but when I try to freeze my model, there is no prediction/truediv node.

I switched it to prediction/Softmax and I am able to save my model, but running it through inference gives me slightly incorrect segmentation. When using the warehouse environment to test, the lanelines come up as yellow obstacles and not as red, but the floor is properly segmented out.

I’m wondering what the div operation is doing, as it seems it’s just a python element-wise divide in the TF docs.

Any help would be appreciated.