NN2SQL: Let SQL Think for Neural Networks

clone the repository

$ git clone https://gitlab.db.in.tum.de/MaxEmanuel/nn2sql

Quickstart with docker

$ docker build . nn2sql
$ mkdir <pathforfigures>
$ docker run -v "<pathforfigures>":/figures nn2sql

Quickstart without docker

$ ./prerun.sh         # extracts data + binaries
$ ./run_benchmarks.sh # runs the benchmarks

Reproduce also long-running SQL-92 queries in Umbra

(un)comment in run_benchmarks.sh:

#export MNISTATTSS="20"
export MNISTATTSS="20 200"
#export MNISTLIMITS="200 2000"
export MNISTLIMITS="2 20 200 2000"

Slowstart

Preparing

$ sudo apt-get install python3 python3-pip r-base
$ pip3 install numpy
$ tar -xf umbra.tar.xz
$ cd mnist && tar -xzf mnist_train.csv.tgz

Scaling the Number of Input Tuples

$ cd iris
$ python3 iris_bench.py > numpy_nn.csv
$ ./iris_sql92_bench.sh | ../umbra/bin/sql
$ ./iris_sql92_psql_bench.sh | psql
$ Rscript iris.r

Image Classification

$ cd mnist
$ python3 mnist_bench.py > numpy_mnist.csv
$ ./mnist_sql92_bench.sh | ../umbra/bin/sql
$ ../umbra/bin/sql < mnist_bench.sql
$ Rscript mnist.r