[scared2]

The main problem is that Tesla autopilot uses pytorch which is a prototype and toying library designed by an intern as compared to tensor flow which was designed from ground up targeting realistic applications such as this one.

[MontyCarloHall]

I'd be very curious to hear some elaboration on this claim; it's the first time I've ever seen it.

From my point-of-view, the fact that TensorFlow completely scrapped its entire ≤v1.0 graph-based API for a PyTorch-style eager execution API for v2.0 onwards does not suggest that it was “designed from the ground up” in any way, shape, or form. We're talking a total rewrite of the entire frontend codebase going from v1.0 to v2.0, and a total abandonment of all v1.0 style code. TF 2.0 is basically a brand new language.

Under the hood it's all the same BLAS/LAPACK operations anyway. I've found TF superior to PyTorch for stats methods and probabilistic programming in general (far richer API), but for standard linear algebra the two are equivalent performancewise, with PyTorch having the superior API IMHO.

[scared2]

Just being easy to use doesn't mean it has a solid back end. I agree pytorch is easier to comprehend UI for average person.

[orbifold]

Completely irrelevant since the model is run on a custom chip. So no part of the pytorch code actually ever makes it into production.

[aeternum]

Comma.ai recently switched their entire codebase from tensor flow to pytorch, how do you explain that?