Donkeycar is a platform to use machine learning racing around tracks.
I am having some trouble with having two labels(angle and throttle) as an input to fastai. I have had a look at MultiCategoryList and was trying to work out how to force the output to be two results, or modifying the CategoryList to return two results?

The input is an image and the result has to be a throttle amount and category.
Both of these labels are binned not continuous so we can treat it as a categorical problem.

So I have been looking at this more with fastai2 and I am trying to work out how to write a new Categorize to allow it to have two Categorical out out thats not MultiCategorize as both output are required to be output. It could also be a two continuous outputs.

The code for Donkeycar bins the inputs and outputs to create categorical model.

I have been looking at this more and you are correct about being a regression model.
Thanks so much for posting your implementation. I will post my implementation of this soon.

I have been thinking of implementing a U-net model with image with image segmentation to get the track geometry. The determining the steering angle and throttle based on the geometry in procedural way.

I was thinking of using this paper [https://arxiv.org/pdf/1807.01726.pdf](LaneNet: Real-Time Lane Detection Networks for Autonomous Driving) but only using the Unet as we only have one lane.

Ok so I ran many different model structures against all my data, this is from many track days. They all seem to over fit quite quickly except for the resnet18. I will look at building my own resnet type model to see if I can make a smaller model and still get good results.
I have added more augmentation as well. I might look at create shadows.