TorchScript and PyTorch JIT | Deep Dive

In this talk, we dive deep into TorchScript and PyTorch JIT.

[EsdrasSoutoCosta]

Amazing! In my graduate research I use PyTorch due to its flexibility and its more close to my thoughts when writing code in order to validate some idea. All of this came from it being a tool fist-based in scientific research and programming experience. PyTorch became the new numpy, but comes with all of this ecosystem that now I can compile my research ideas to deploy it into production ready server. Awesome!

[keinemoto]

I really love PyTorch. It's flexible and has great supports for production.

[pixel]

Wish we had more people involved on the C++ endpoint deployment for Libtorch #drogon #rustBinding #golangBinding

[ml-cfd]

Great video! Plenty of useful insights. I would like to see more of these videos :-)
Cheers

[vascodev]

Thanks for the great video. I was only wondering how the last part about PyTorch JIT compilation for more dynamic setups, differs from the scripting (or tracing) part explained earlier in the video. Is this an extra step we can take alongside scripting? I am kind of confused since both scripting and tracing are part of the jit module, and since scripting (I believe) already allows for more dynamic models as explained in the video, so I am trying to figure out how this differs. Thanks in advance for any clarification!

[midhileshelavazhagan]

Why there is not fuse group present in the IR representation? How can we visualize the optimization made by torchscript over the model?

[MadlipzMarathi]

Will torchscript allow run models on opencl backend at runtime

[ArunRangarajan]

Is James a machine? :D (Like Commander Data from Starship Enterprise?)

[clivefernandes]

Hi can anyone suggest a place or book where I can learn more about these optimizations

[orjihvy]

How about Torchscript into Julia instead of C++?

[jonathansum]

So I can write Bert, transformer, and more advance model in Pytorch and run them in Java and C++?