Sie sind auf Seite 1von 2

Features Business Explore Pricing This repository Search Sign in or Sign up

juliandewit / kaggle_criteo Watch 4 Star 32 Fork 16

Code Issues 0 Pull requests 0 Projects 0 Pulse Graphs

Software for the kaggle criteo challenge

20 commits 1 branch 0 releases 1 contributor MIT

Branch: master New pull request Find file Clone or download

juliandewit Rename Latest commit 07ece28 on 6 Oct 2014

Criteo Rename 3 years ago

Lib First commit 3 years ago

SharpNet Finalization 3 years ago

.gitignore 64bit 3 years ago

Criteo.sln First commit 3 years ago

LICENCE.md Update LICENCE.md 3 years ago

Model documentation.docx Finalization 3 years ago

ModelDocumentation.pdf Rename 3 years ago

README.md Update README.md 3 years ago

README.md

Kaggle Criteo
This is the software for the kaggle criteo challenge that ended up at the 4th place. The algorithm is a GPU based "deep"
neural network. Multiple models were trained using varied types of preprocessing and network architectures. The models
were averaged into one final model.

The software

The software is written in c# .Net framework 4.5 using VS 2012. GPU programming is done using Cudafy.NET library. CUDA
5.X

The neural network is an ongoing project of mine and I'm still shaping and moving and renaming the code to make it sensible
for me. The version in this solution is a subset of all the functionalities. Only the moving parts were put in this repository.

Compiling

If you want to compile the project you should install the NVIDIA Cuda SDK first. Next to this you need VS2012. VS2013 might
work. Although not completely necessary you can install the Cudafy.NET library. With the cudafy tools you can check if
everything is in working order. If so, compiling the project should not give major difficulties.

Running

First of all you need a 64bit windows OS. The project was done on a workstation with 32Gb memory. Perhaps it will still run
with a little bit less. Most likely 16Gb is not enough. The problem is that training is done with the complete dataset in memory
for high throughput. Last but not least you need a recent GPU board with cuda compute capability 3.
2017 GitHub, Inc. Terms Privacy Security Status Help Contact GitHub API Training Shop Blog About

Das könnte Ihnen auch gefallen