SystemDS-NN Examples

This folder contains scripts and PySpark Jupyter notebooks serving as examples of using the SystemDS-NN (nn) deep learning library.

Examples

MNIST Softmax Classifier

• This example trains a softmax classifier, which is essentially a multi-class logistic regression model, on the MNIST data. The model will be trained on the training images, validated on the validation images, and tested for final performance metrics on the test images.
• Notebook: Example - MNIST Softmax Classifier.ipynb.
• DML Functions: mnist_softmax.dml
• Training script: mnist_softmax-train.dml
• Prediction script: mnist_softmax-predict.dml

MNIST “LeNet” Neural Net

• This example trains a neural network on the MNIST data using a “LeNet” architecture. The model will be trained on the training images, validated on the validation images, and tested for final performance metrics on the test images.
• Notebook: Example - MNIST LeNet.ipynb.
• DML Functions: mnist_lenet.dml
• Training script: mnist_lenet-train.dml
• Prediction script: mnist_lenet-predict.dml

Neural Collaborative Filtering

• This example trains a neural network on the MovieLens data set using the concept of Neural Collaborative Filtering (NCF) that is aimed at approaching recommendation problems using deep neural networks as opposed to common matrix factorization approaches.
• As in the original paper, the targets are binary and only indicate whether a user has rated a movie or not. This makes the recommendation problem harder than working with the values of the ratings, but interaction data is in practice easier to collect.
• MovieLens only provides positive interactions in form of ratings. We therefore randomly sample negative interactions as suggested by the original paper.
• The implementation works with a fixed layer architecture with two embedding layers at the beginning for users and items, three dense layers with ReLu activations in the middle and a sigmoid activation for the final classification.

Setup

Code

• To run the examples, please first download and unzip the project via GitHub using the “Clone or download” button on the homepage of the project, or via the following commands:

  git clone https://github.com/dusenberrymw/systemml-nn.git
  

• Then, move into the systemml-nn folder via:

  cd systemml-nn
  

Data

• These examples use the classic MNIST dataset, which contains labeled 28x28 pixel images of handwritten digits in the range of 0-9. There are 60,000 training images, and 10,000 testing images. Of the 60,000 training images, 5,000 will be used as validation images.
• Download:
  • Notebooks: The data will be automatically downloaded as a step in either of the example notebooks.
  • Training scripts: Please run get_mnist_data.sh to download the data separately.

Execution

• These examples contain scripts written in SystemDS's R-like language (*.dml), as well as PySpark Jupyter notebooks (*.ipynb). The scripts contain the math for the algorithms, enclosed in functions, and the notebooks serve as full, end-to-end examples of reading in data, training models using the functions within the scripts, and evaluating final performance.

• Notebooks: To run the notebook examples, please install the SystemDS Python package with pip install systemds, and then startup Jupyter in the following manner from this directory (or for more information, please see this great blog post):

  PYSPARK_DRIVER_PYTHON=jupyter PYSPARK_DRIVER_PYTHON_OPTS="notebook" pyspark --master local[*] --driver-memory 3G --driver-class-path SystemDS.jar --jars SystemDS.jar
  

  Note that all printed output, such as training statistics, from the SystemDS scripts will be sent to the terminal in which Jupyter was started (for now...).

• Scripts: To run the scripts from the command line using spark-submit, please see the comments located at the top of the -train and -predict scripts.