Vectorflow: Minimalist neural network library faster than TensorFlow in D | Hacker News Reader

Vectorflow: Minimalist neural network library faster than TensorFlow in D

github.com

85 pointsteleforce17 comments

I wonder if "faster" really matters. The implementation-level optimization seems paled when compared to productivity, ecosystem, and algorithm-level and hardware-level boost. Case in point, Caffe2 wanted to be the backend of PyTorch in FB, yet failed. MxNet claims to be "faster" than PyTorch in pretty much every way and Smola personally pushed orgs in AWS AI really hard to adopt MxNet and Gluon* yet failed. The latter failure is particularly worth mentioning: all teams in AWS AI complied and adopted MxNet but then three years later decided to move to PyTorch. All-in-all, the incremental speed boost was simply irrelevant.

I keep a few projects at work where we train models with Keras+TensorFlow. Twice before I tried to switch to PyTorch, because it did have nicer features/syntax for research and SOTA models, but deploy and support was limited on the toolchain we use, so we had to stay with Keras+TF. I'd still like to switch to PyTorch because of a few goodies, but the deploy of the models on some platforms is still simpler starting from a TF model.

If anyone has a few links or resources to point me to, I'm more than happy to make a third try.

if there was constant deployment in some large industrial niche, faster implementations might help you save on your electric bill

This is interesting.

Modern NN implementations (Pytorch/TF) are optimised for training deep neural networks on GPUs.

This is instead designed for shallow neural networks on CPUs. That's similar to the original Word2Vec implementation, and there are lots of similar use cases where this could be useful

I am curious to know more about what lead to D lang being chosen for this project?

Check the blog post

- They wanted the same language for implementation and user API

- They wanted a single machine, CPU bound runtime where the executable is the model

Given these, D and Go make sense - easy to pick up, compiles fast and performant.

Thanks for heads up on blog had missed that. Great!

D I understand, but Go is not a particularly performant language.

Why is D so unpopular? It seems to be a decent language.

Not enough marketing. Too much competition in the alternative systems language space.

Vectorflow was my gateway drug to Dlang.

what does it take to get a good word-to-vec--style model going- or are those not 'sparse' ?

They should be sparse -- the input to Word2Vec is one-hot encodings of words in the models dictionary

Is it faster though when it's running on a distributed system with 128 machines?

“ Vectorflow is a minimalist neural network library optimized for sparse data and single machine environments.”

What are interesting things you can do with a minimalist neural network, or even a full blown neural network for that matter?

You just kind of like... train a data set, hope you don't under or overtrain, and then what? Feed it some input, get out an approximate output?

Their usecase is explained in the accompanying blogpost. https://netflixtechblog.medium.com/introducing-vectorflow-fe...

Reply on news.ycombinator.com