Search results
View Andrew Tulloch’s profile on LinkedIn, a professional community of 1 billion members. Experience: OpenAI · Education: University of Cambridge · Location: Menlo Park · 500+ connections on...
- 500+
- 1.1K
- OpenAI
- San Francisco Bay Area
Oct 3, 2021 · Andrew Tulloch — Machine Learning, Statistics, Systems About | Academic | GitHub | CV Improving PyTorch inference performance on GPUs with a few simple tricks
199. 2018. Software-hardware co-design for fast and scalable training of deep learning recommendation models. D Mudigere, Y Hao, J Huang, Z Jia, A Tulloch, S Sridharan, X Liu, ... Proceedings of the 49th Annual International Symposium on Computer …. , 2022.
Apr 10, 2012 · I'm Andrew Tulloch. I'm an engineer at Facebook. I'm available on GitHub, LinkedIn, Twitter, and Facebook.
- Naive Method
- Flattened Tree Method
- Compiled Tree Method
- Performance Evaluation
- Conclusions
Superficially, decision treeevaluation is fairly simple - given afeature vector, recursively walk down the tree, using the givenfeature vector to choose whether to proceed down the left branch orthe right branch at each point. When we reach a leaf, we just returnthe value at the leaf. In Haskell, In C++, From now on, we'll focus on the C++ implemen...
A nice trick to improve cache locality is to lay out data out in aflattened form, and jumping in between locations in our flattenedrepresentation. This is analogous torepresenting a binary heap as an array. The technique is just to flatten the tree out, and so moving from theparent to the child in our child will often mean accessing memory inthe sa...
A really cool technique that has been known for years is generating Ccode representing a decision tree, compiling it into a shared library,and then loading the compiled decision tree function via dlopen(3).I found a 2010 UWash student reportdescribing this technique, though the earliest reference I've seen isfrom approximately 2000 in a presentatio...
As the student reportindicates, therelative performance of each strategy depends on the size of thetrees, the number of trees, and the number of features in the givenfeature vector. Our methodology is to generate a random ensemble with a given depth,number of trees, and number of features, construct the evaluators ofthis tree for each strategy, and...
We've implemented and analyzed the performance of a selection ofdecision tree evaluation strategies. It appears there are two mainconclusions: 1. For small models - <200 or so trees with average depth <2, thecompiled evaluation strategy is the fastest. 2. For larger models, the piecewise flattened evaluation strategy ismost likely the fastest. 3. C...
Andrew George Tulloch (born 1 April 1967 in Wolverhampton [1]) is a male retired English athlete who specialised in the 110 metres hurdles. [2] Athletics career. He represented Great Britain at the 1996 Summer Olympics, as well as three consecutive World Championships, starting in 1993.
Improving PyTorch inference performance on GPUs with a few simple tricks —Andrew Tulloch. 3 October 2021. We hear a lot these days that "machine learning systems are stuck in a rut".
