I just tried this for the sake of playing around. I was curious at how the Resnet34 model will perform given how similar the Model3 and ModelS cars are. I was quite surprised at the results :).
I ignored all the download errors related to content size and all that. Goes to show how robust the model is. I tried to increase the fit_cycle to 8 but started running into CUDA OOM errors. I am using GCP. Anyways great experience. One mistake I noticed is that I failed to change the learning rate interval. I tried to correct this but am landing with CUDA OOM errors. Will keep trying and share the results.
I reused the download_images.ipynb notebook but cleaned up a bit. My results: