G-SLIDE

G-SLIDE is a GPU-based sub-linear deep learning engine via LSH sparsification of fully-connected neural networks. The details can be found in this paper.

Dataset

The Datasets can be downloaded in Amazon-670K and WikiLSHTC-325K.

Baselines

The baseline is SLIDE. The source codes of TensorFlow-CPU and TensorFlow-GPU baselines can also be found from the same link.

Running G-SLIDE

Tested Environments

The experiment environments in the paper are as follow:

OS: Ubuntu 20.04
Compiler: nvcc 11.1
GPU: 2080ti
CPU: Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz
CMake：3.14 and above

Dependencies

cuBLAS
Thrust
JsonCpp: we use it to parse the configuration json file.

Compile and Run

Type the following commands to compile the project:

git clone https://github.com/PanZaifeng/G-SLIDE.git
cd G-SLIDE
cmake -B build
cmake --build build

Before running G-SLIDE, you should download the dataset of Amazon-670K and re-configure the amazon.json properly. Note that there will be lots of information to be printed, so we recommend redirecting stdout when running.

./runme ./amazon.json > amazon.log