TimoHackel / ILA-SCNN

Computation using data flow graphs for scalable machine learning

---

ILA-SCNN

This repository contains the source code of our paper: Inference, Learning and Attention Mechanisms that Exploit and Preserve Sparsity in Convolutional Networks Our implementation extents tensorflow by providing GPU implementations for layers that are able efficiently process large-scale, sparse data. For more details please refer to our paper.

Installation

After cloning this repository:

1. Follow: Installing TensorFlow to install tensorflow from source code.

2. build our sparse modules:

bazel build -c opt --config=cuda --strip=never -s //tensorflow/core/user_ops:direct_sparse_conv_kd.so

3. set environment variables

echo 'export LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:$HOME/src/ILA-SCNN/bazel-bin/tensorflow/core/user_ops"' >> ~/.bashrc
echo 'export PYTHONPATH="$PYTHONPATH:$HOME/src/ILA-SCNN/tensorflow/core/user_ops"' >> ~/.bashrc
source ~/.bashrc

Examples

We provide the following examples:

1. Modelnet Application
2. MNIST Application

The correct data sets will be downloaded automatically when running the scripts.