The Evolution of Deep Learning and PyTorch

  • by
The Evolution of Deep Learning and PyTorch

The evolution of deep learning and PyTorchPyTorch is an open source, machine learning framework used for writing Python programs that describe an algorithm that can do artificial intelligence. People use PyTorch to do fundamental AI research so that we build better building blocks that can, that you can build applications on top of.

And then people use these building blocks to build more advanced AI models in specific fields. For example, for better self-driving cars, for better speech recognition. The field of AI, in middle to late 2016, was grappling with the idea of doing neuro-networks that were more and more dynamic. The other problem was that there were rigidity in the networks you could define.

Researchers wanted to make them more determined by the network at run time rather than predefined beforehand. And that’s what tools before, that were existing didn’t really give a good way to do so and that’s roughly what PyTorch provided and that’s one of the big reasons why it exploded. It gives you a deep learning framework that brings dynamic neuro-networks and input to programming as first class citizens.

When people build these applications, they want to actually deploy them at scale. We basically created a subset of Python called TorchScript which is much more easily deployable. What we give you is a seamless workflow where you start with research and then you start adding function annotations to your code, the part of the code that you actually want to ship to production.

But in this subset of Python, if you annotate your code to say, my code is TorchScript compatible and it can be run by a Torch.jit compiler, then our compiler takes the code, it analyzes it, it modifies it and it will run it faster and it can export it into our own intermediate representation and you roughly get all the benefits of a compilation process, that you would expect.

We think this is one of the big, future directions for just not PyTorch but also the field. Wherever the field goes, we probably will have to adapt, either with modifying PyTorch as it is today or probably, at some point, rewriting PyTorch plus plus.


Useful links:

reference – The evolution of deep learning and PyTorch

Share this post ...

Leave a Reply

Your email address will not be published. Required fields are marked *