61–70 of 98 posts
Why would we want to use purely functional data structures? When do the pros of functional data structures outweigh the additional complexity? Are there scenarios when a project would want to pivot from a regular data structure to a purely functional one?
Why would we want to use purely functional data structures? When do the pros of functional data structures outweigh the additional complexity? Are there scenarios when a project would want to pivot from a regular data structure to a purely functional one?
You can also safely re-use sub-structures without performing a deep copy. For example, if you want to keep a sub-tree around for later you can do that in O(1) time because it's safe to keep a reference to it. If it is a mutable tree you don't know what's going to happen to it so you need to do a deep copy of the entire sub-tree you're holding on to. This can save a lot on memory allocation and copying depending on your use case.
Why would we want to use purely functional data structures? When do the pros of functional data structures outweigh the additional complexity? Are there scenarios when a project would want to pivot from a regular data structure to a purely functional one?
There are obviously other trade offs as well, like performance and memory usage.
Why would we want to use purely functional data structures? When do the pros of functional data structures outweigh the additional complexity? Are there scenarios when a project would want to pivot from a regular data structure to a purely functional one?
Traditionally, OO code is written all the time.
But after I learned JAX/Flax, a light turned on inside my head and I now write functional Deep Learning code as much as I can. All my side projects and new code are purely functional in JAX/Flax. PyTorch has functional API known as functorch, and I have used it one project.
Where lots and lots of data in 3,4,5 dimensional tensors exist, and you need to run lots of transformation on them, and then you need to multiply a huge number of them thousand times in each second- functional code makes much more sense and gives immense sanity and peace of mind.
Those of you writing Deep Learning code, learn functional programming principles (immutable data, pure functions, leaving no side effect, etc.), and apply them to DL via functorch or JAX.
Your life will never be the same.
I'm reading this book right now. It's really great so far! I've been working a lot with Trees in Clojure, and have been hitting serious limitations of my understanding. I also found this YouTube video from a Clojure conference that reviews some different strategies for tree traversal in Clojure: https://youtu.be/YgvJqWiyMRY I thought that learning a Functional Lisp would make it really easy to traverse trees, since t…
I feel like they are.. not so awesome: they are grossly inefficient due to all the pointer chasing and are pretty much guaranteed to be slower than the alternative since they trash your cache.
One of the key points of immutable datastructures is that they're easy to reason about given that the invariants are upheld. Designing an efficient memory representation that upholds those invariants is an entirely separate problem domain. Not every pointer dereference is slow, and not every tree is represented with pointers.
It's weird that there are so many claims in here that the data structures and algorithms are perfectly performant yet there isn't even one look at generated assembly or any acknowledgement of the underlying system that is supposed to run the code. Proving things are Big O performant is neat, but at some point the code has to hit hardware.
Why would we want to use purely functional data structures? When do the pros of functional data structures outweigh the additional complexity? Are there scenarios when a project would want to pivot from a regular data structure to a purely functional one?
Deep Learning applications is one area. Traditionally, OO code is written all the time. But after I learned JAX/Flax, a light turned on inside my head and I now write functional Deep Learning code as much as I can. All my side projects and new code are purely functional in JAX/Flax. PyTorch has functional API known as functorch, and I have used it one project. Where lots and lots of data in 3,4,5 dimensional tensors…
Where to learn about it other than the documentation?
It's weird that there are so many claims in here that the data structures and algorithms are perfectly performant yet there isn't even one look at generated assembly or any acknowledgement of the underlying system that is supposed to run the code. Proving things are Big O performant is neat, but at some point the code has to hit hardware.
Yeah, Big-O notation is only part of the picture and yes, "galactic algorithms" that are theoretically efficient but only for astronomically big inputs do exist, but in many cases (especially if your algorithm isn't doing anything particularly deep or clever) linear is faster than quadratic which is faster than exponential on real world input, too. Mathematics is the science of modeling and abstraction and that abstr…
Earlier quoted context omitted.
Deep Learning applications is one area. Traditionally, OO code is written all the time. But after I learned JAX/Flax, a light turned on inside my head and I now write functional Deep Learning code as much as I can. All my side projects and new code are purely functional in JAX/Flax. PyTorch has functional API known as functorch, and I have used it one project. Where lots and lots of data in 3,4,5 dimensional tensors…
Do JAX and functorch have the same level of builtin functionalities (operations) as the original Pytorch library? Where to learn about it other than the documentation?
Functorch is still new, and honestly, there is little to learn if you already know JAX. There are some talks from Meta, and then there is always the docs.
I'm reading this book right now. It's really great so far! I've been working a lot with Trees in Clojure, and have been hitting serious limitations of my understanding. I also found this YouTube video from a Clojure conference that reviews some different strategies for tree traversal in Clojure: https://youtu.be/YgvJqWiyMRY I thought that learning a Functional Lisp would make it really easy to traverse trees, since t…
I feel like they are.. not so awesome: they are grossly inefficient due to all the pointer chasing and are pretty much guaranteed to be slower than the alternative since they trash your cache.
You might not be running GHC's RTS on a resource-constrained target platform but you can generate the C code that could run on that platform from Haskell, leveraging the guarantees you get with using Haskell's advanced type system.
Update: point is that GHC can optimize code and some of those optimizations can avoid pointer chasing! But like most optimizing compilers, some times GHC can miss them, and you have to do a bit of work to tell GHC where it can avoid chasing unnecessary pointers (or where it should/should-not inline, give it some help to know where it can do fusion, etc).