Hacker News Reader

Comment by TheAnig | Hacker News Reader

Sound like a pretty neat introduction! This is exactly the kind of thing I needed, coming from tf/keras and looking to switch to pytorch for research

View on HN

As someone learning Keras right now, why are you wanting too switch?

Not the parent, but the imperative interface supported by the dynamic graph approach Pytorch takes is much nicer.

Additionally, in my personal opinion Tensorflow is often too low level and Keras is often too high level for the things I'm trying to do for research. While you can jump between the two of course, I think PyTorch hits a much more natural middle ground in its API.

Tensorflow/Keras is making improvements in these areas with the eager execution, and is still great for putting models into production, but I think PyTorch is much better for doing research or toying with new concepts.

This article has some good comparison: http://www.goldsborough.me/ml/ai/python/2018/02/04/20-17-20-...

Also, I believe the newer fast.ai course uses Pytorch, and the prior used Keras?

I would like to know more from the article about setting x,y,m1, and m2. Any explanation is appreciated.

any reason why you wouldn't use tf eager?

tf.eager is much slower than pytorch.

this is a strong claim. what is your source for this? I'm legitimately curious.

No hard benchmarks, just personal experience. Note that I’m not saying regular Tensorflow is slower than pytorch (in fact I’ve found them to be roughly the same) just eager mode.

Edit: Just realized this might be a good thing to write a blog post about. I’ll get back to you after finals :)

I'm genuinely interested if you ever do write that blog post

Update: other people have already scooped me:

https://medium.com/@yaroslavvb/tensorflow-meets-pytorch-with...

I've seen similar performance regressions on my own tasks and I don't have much to add beyond what's in that blog post.

Reply on news.ycombinator.com