Why AI and machine learning researchers are beginning to embrace PyTorch | Hacker News Reader

Why AI and machine learning researchers are beginning to embrace PyTorch

oreilly.com

54 pointsrfreytag12 comments

I haven't observed anything related to what the article is talked about. I haven't heard of any hype related to PyTorch at all other than the announcements by Facebook. Can someone correct me if I am wrong?

I've seen a recent uptick in the number of people using PyTorch in the research community, but it's still a very small proportion of the total and likely won't be enough to overcome TensorFlow's network effect.

As a data nerd not doing deep learning but who follows the space, I'd say the article jives with what I've seen and heard: pytorch is cool, but predominantly for research and prototyping. Both academic and non-academic colleagues use it to test ideas a bit more quickly before taking the time to write it in tensorflow for larger-scale training. I suspect as the distributed stuff new to 0.2 catches on, the more academic ones might be able to stick with pytorch for the full/distributed training runs; curious to see.

So if it's good for R&D now, it should disrupt its way up the value chain over time, right?

Research and experimentation with novel architectures has switched to pytorch everywhere I look. All colleagues and myself outside production pipelines have switched to pytorch. TF is good but too messy and hard for R&D, that's the price to pay for using a production-grade tool in research operations.

Agree. Everyone I interact with it is all TF. Even my son at University they are using TensorFlow. Hype or simply better at this point does not really matter as it is simply self fulfilling at this point.

Think of it like k8s and Docker as the same.

PyTorch v0.2.0 was just released. The HN thread and related discussion is here: https://news.ycombinator.com/item?id=14946585

I LOVE PyTorch for experimenting with dynamic deep neural nets (DNNs) -- that is, DNNs that can have different graphs for different input samples. I find it much, MUCH easier to create and tinker with dynamic DNNs using PyTorch than, say, TensorFlow Fold.

Pytorch is awesome especially with automatic differentiation. We used to do the training in torch and then convert the trained model to caffe. The only problem is that the conversion script supports only torch 7. I heard that at Facebook they experiment and do the training in pytorch and then they convert the model to caffe 2. I hope they will release the conversion script.

Let's keep in mind that computer book publishers (like O'Reilly) have an interest in thrashing the knowledge base of programmers generally. Because it means more potential book sales.

Advertisers can be informative. But they can also be a big distraction of time & money & resources.

So, on one hand, the "pick an OS/Python version/package manager" selector for getting install instructions is cool and useful.

On the other hand, well, having installers for OS X and Linux leaves a big Windows-sized gap, so I won't be playing with this anytime soon.

pytorch is my go to framework for deep learning. That given this is just a PR piece, an interview from Oreilly with one of the main developers of pytorch.

How does pytorch compare to dynet (formerly CNN, or pyCNN)?

It's been my choice for dynamic NN for a while.

Reply on news.ycombinator.com