Skip to content

Latest commit

 

History

History
58 lines (37 loc) · 1.76 KB

README.md

File metadata and controls

58 lines (37 loc) · 1.76 KB
Raw

Fiddling with monitor data time series

AR-204 on JIRA. Learning from time series to predict future, in short and long range.

Notebooks

Notebooks are in the notebooks/ folder.

  • sandbox.ipynb --- ignore this.
  • monitor-data.ipynb --- extracting raw monitor data from the dataset.
  • visualize-predicts.ipynb --- visualization of predictions for a single stay.
  • predict-stats.ipynb --- prediction performance evaluation.

Comand-line utilities

Run <command> -h to get help on command-line options.

  • extract --- extracts contiguous fragments from data for training
  • prepare --- prepares the dataset as a tensor for training
  • train --- trains a model on the dataset
  • predict --- extends a stay with predictions and log-likelihoods of observations.

Running experiments

Overview

The stays are in data/. The order of preparing a data set for training is

  • extract
  • prepare

Training

Use train to train a model. The depth must be greater than 1 for the model to learn to predict more than a single step in the future and deal with missing values.

Prediction

Use predict to extend a frame with prediction. Models are in models/, data folder should be a subdirectiory of data/ (with the list of concept names as the name) and the stay is a pickle file in data/. The model and the data must match. Example:

predict model/ALL-15.model data/ALL data/monitor-dataset-Ichilov_MICU_20194.pkl

Helper scripts

  • scripts/extract.sh --- extracts contiguous fragments (120 min by default).
  • scripts/train.sh --- trains a model.

For scripts, modify defaults by specifying environment variables on the command line. For example:

RATE=0.0001 scripts/train.sh

to set the training rate to 0.0001. Consult the scripts for available variables.