Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Deep Learning with PyTorch

pytorch.org

Post reply on HN
Print view View original

21–30 of 40 posts

Re: Deep Learning with PyTorch

#21

Post by suresk »

earthnail wrote: post #16
madenine wrote: post #10

Earlier quoted context omitted.

Goodfellow’s book on deep learning[0] is a good starter - the first chapters give a solid overview of ML theory as well. Elements of Statistical Learning is another. [0] http://www.deeplearningbook.org/

I don’t think it is a good book. From a didactic point of view, I actually found it one of the worst resources out there. The math intro at the beginning is too superficial - either you know it and skip the chapter, or you need another resource to learn. The rest of the first part is okay, but parts 2 and 3 are really not very helpful to someone who doesn’t already understand it. I strongly recommend fast.ai instead.…

I wasn't sure if I was just dumb or if everyone recommending this book just kept it on their shelves and talked about it, sort of like Knuth or CLRS. I don't necessarily think it is a bad book, but it is hard to think of a better one to give someone to discourage them from getting into deep learning.

I started with a top-down approach via the fast.ai courses and learning Keras, then spent time brushing up on some of the math concepts (as you said, it assumes a fair amount of previous knowledge), and then went back and started re-reading it, and I finally feel like I'm starting to get some value out of it.

Definitely wouldn't recommend it as a first book, though.

Re: Deep Learning with PyTorch

#24

Post by t-vi »

Mirioron wrote: post #9

What kind of knowledge should you have before getting into this book? I've been meaning to try to learn ML and have been looking at some university courses that have all the material available.

The idea is that you should know basic maths (say what matrix multiplication does and what a vector space is) and a bit of Python. We work very hard on keeping everything hands-on from there. (Disclaimer: I'm not unbiased.)

Re: Deep Learning with PyTorch

#25

Post by bitL »

mroche wrote: post #12

For those in the ML area, what are people’s opinions on PyTorch and its use in comparison to its competition? I don’t have any experience with PyTorch or ML tech besides having to package and provide PyTorch containers for our university’s HPC cluster and running the helloworld.py against it for validation.

PyTorch is research favorite, it's easy to debug tensors and interleave bunch of random code in between layers if you need to. Keras is the easiest one to get a model running in production, but for some advanced things like sophisticated custom losses you might be forced to switch over to base TensorFlow. TensorFlow is the production king but it's super complex unless you use tf.keras and difficult to debug.

Re: Deep Learning with PyTorch

#26

Post by e19293001 »

Mirioron wrote: post #9

What kind of knowledge should you have before getting into this book? I've been meaning to try to learn ML and have been looking at some university courses that have all the material available.

Grokking Deep Learning - while it teaches the very basics of deep learning, it claims to teach you basic PyTorch at the end of the book.

https://www.manning.com/books/grokking-deep-learning

Re: Deep Learning with PyTorch

#28

Post by sseveran »

mroche wrote: post #12

For those in the ML area, what are people’s opinions on PyTorch and its use in comparison to its competition? I don’t have any experience with PyTorch or ML tech besides having to package and provide PyTorch containers for our university’s HPC cluster and running the helloworld.py against it for validation.

So we have been a TensorFlow shop since well before 1.0. We are keeping tabs on PyTorch with an eye on moving in that direction as certain things get filled, notably serving and something that is really comparable to tf.data.

The fact that our team in particular is looking at moving is probably a disaster for TF since a number of the criticisms typically leveled at it we didn't see as issues. We have a couple of really strong people with haskell backgrounds so static graphs and laziness we didn't find to be problematic.

However the embrace of Keras in 2.0 has left us dumbfounded. On one hand having a consistent layer interface is nice. On the other hand having a base class for the loss function that is not sufficiently general, the fact that all non-toy models we build seem to need model subclassing and a custom training loop with GradientTape and the number of issues we ran into while trying to port a couple of models has led me to conclude that the release was not ready. So while we like the tools around the model (tf.data, tensorboard, serving, tfx, etc...) building actual models I think has gotten worse.

Now my opinions on PyTorch are not from shipping production models but mostly porting to TF and keeping tabs on what they are doing. PyTorch also makes it easy to define reusable units. It does not try and expose a higher level interface that requires a significant investment in learning to express complex or unusual models. It seems a bit less opinionated on what the user should do.

A couple of other notes, PyTorch is being used inside Google for research I think. They have now written several papers (including one with Jeff Dean as 5th author) that have had their code released in PyTorch. PyTorch I think (it might already have) will end up with better governance but I would be interested in others opinions. They have at least one person listed under the project maintainers who does not work for FaceBook. A reason for adopting PyTorch may be that one company just does not decide to radically change the project to fit their view of the world.

This last bit is purely conjecture. PyTorch I think has already won over TF and it is going to take a couple of years for it to play out. If I had to bet today I would bet that PyTorch will become the dominate framework for both research and production. Of course something could happen to derail that but if things continue on their current trajectories I think its inevitable.

Re: Deep Learning with PyTorch

#30

Post by acollins1331 »

sseveran wrote: post #28
mroche wrote: post #12

For those in the ML area, what are people’s opinions on PyTorch and its use in comparison to its competition? I don’t have any experience with PyTorch or ML tech besides having to package and provide PyTorch containers for our university’s HPC cluster and running the helloworld.py against it for validation.

So we have been a TensorFlow shop since well before 1.0. We are keeping tabs on PyTorch with an eye on moving in that direction as certain things get filled, notably serving and something that is really comparable to tf.data. The fact that our team in particular is looking at moving is probably a disaster for TF since a number of the criticisms typically leveled at it we didn't see as issues. We have a couple of real…

I echo this. While 2.0 initially made me happy building some models from scratch, as soon as I needed to write my own optimizers, losses, class weights, etc. it became a nightmare. Also not to mention the data pipeline for large imagesets required you to serialize the data first into tfrecords. Please just stream AND shuffle images in a folder tensorflow.
Post reply on HN