hastl 0.1.8

A fast GPU implementation of STL decomposition with missing values

HaSTL [ˈheɪstiɛl]: A fast GPU implementation of batched Seasonal and Trend decomposition using Loess (STL) [1] with missing values and support for both CUDA and OpenCL (C and multicore backends are also available). Loosely based on stlplus, a popular library for the R programming language. The GPU code is written in Futhark, a functional language that compiles to efficient parallel code.

Requirements

You would need a working OpenCL or CUDA installation/header files, C compiler and these Python packages:

• futhark-ffi==0.14.2

• wheel

Installation

You may want to run the program in a Python virtual environment. Create it via:

python -m venv env

Then, activate the virtual environment via:

. env/bin/activate

Upgrade pip via:

pip install --upgrade pip

Then select the backends (choose from opencl, cuda, c and multicore) that you wish to build by setting the environment variable:

export HASTL_BACKENDS="opencl multicore c"

If no environmental variable is set, only the sequential c backend would be compiled.

The package can then be easily installed using pip. This will take a while, since we need to compile the shared libraries for your particular system, Python implementation and all selected backends:

pip install hastl

To install the package from the sources, first get the current stable release via:

git clone https://github.com/mortvest/hastl

Install the dependencies via:

pip install -r requirements.txt

Afterwards, you can install the package. This can also take a while:

python setup.py sdist bdist_wheel
pip install .

Usage

Examples of HaSTL usage can be found in the examples/ direcotry. The simplest snippet should contain:

from hastl import STL
stl = STL(backend=..)
seasonal, trend, remainder = stl.fit(data, n_p=.., q_s=..)

References

[1] Cleveland, Robert B., et al. “STL: A seasonal-trend decomposition.” J. Off. Stat 6.1 (1990): 3-73.