dmartyanov/USI-RNN

Modeling Event Sequences with Unevenly Spaced Intervals with RNN

This project is a demo example of how sequences with unevenly spaced intervals can be modeled for processing via Neural Network

Tensorflow with Keras API is used to train and evaluate models

Prerequisites

• tensorflow r1.4
• keras 2.0.4
• numpy 1.12.1

Getting Started

Installation

• Install tensorflow from https://github.com/tensorflow/tensorflow
• Clone this repo:

git clone https://github.com/dmartyanov/usi-rnn
cd usi-rnn

• Install requirements:

pip install -r requirements.txt

Train

Add your dataset to ./data/

P - padding length for sequences, default is 200

M - number of examples

A - alphabet size, default is 252

• ./data/X.npy - (M, P) numpy array with events padded to PSL with 0 value, max value is less than A
• ./data/Y.npy - (M, ) numpy array with binary target values
• ./data/L.npy - (M, ) numpy array with the lengths of the sequences before padding
• ./data/D.npy - (M, P) numpy array with time labels padded to PSL with 0 value, non-decreasing sequences
• ./data/tr_idx.npy - (TRAIN_LENGTH, ) numpy array with train indices
• ./data/cv_idx.npy - (VALIDATION_LENGTH, ) numpy array with validation indices

Run:

python main.py

References

Slides