openturns / otmorris Goto Github PK

View Code? Open in Web Editor NEW

0.0 4.0 6.0 993 KB

Morris sensitivity method module

Home Page: http://openturns.github.io/otmorris/master/

License: GNU Lesser General Public License v3.0

CMake 27.73% Shell 2.50% Makefile 0.97% NSIS 12.33% C++ 40.00% C 0.44% Python 8.62% TeX 3.74% SWIG 3.68%

morris-method sensitivity openturns

otmorris's Introduction

OTMORRIS Module

OTMORRIS is a module for OpenTURNS

More information can found at http://www.openturns.org

Installation

Please see the http://openturns.github.io/openturns/latest/install.html for instructions on installing OpenTURNS modules on various platforms from binaries or sources.

-- The OpenTURNS team

otmorris's People

Contributors

Watchers

Forkers

jschueller sofianehaddad tian-wei adumasphi mbaudin47 yomgarcia

otmorris's Issues

Example 1 from OTMorris documentation fails

Here is the problematic example: http://openturns.github.io/otmorris/master/examples/examples.html#example-1-morris-use-case-and-p-level-input-grid

The documentation has two problems. The first problem is that this line in the example:

ee = otmorris.Morris(X, Y)

leads to the following errors:

TypeError: Wrong number or type of arguments for overloaded function 'new_Morris'.
  Possible C/C++ prototypes are:
    OTMORRIS::Morris::Morris()
    OTMORRIS::Morris::Morris(OT::Sample const &,OT::Sample const &,OT::Interval const &)
    OTMORRIS::Morris::Morris(OTMORRIS::MorrisExperiment const &,OT::Function const &)
    OTMORRIS::Morris::Morris(OTMORRIS::Morris const &)

This appears to be fixable by changing the line to the following:

bounds = morris_experiment.getBounds()
ee = otmorris.Morris(X, Y, bounds)

The second problem is discussed in issue #26.

Module handle only uniform distributions

The module handle only uniform inputs.
If input model contains non-uniform distributions, there is not any framework to handle them

Documentation for Morris is inconsistent about the available constructors

The documentation in question is here: http://openturns.github.io/otmorris/master/user_manual/_generated/otmorris.Morris.html#otmorris.Morris

The top of the documentation indicates that there are two constructors with the following signatures:

Morris(inputSample, outputSample, interval)
Morris(experiment, model)

However, later text indicates that there are five constructors:

With first constructor, we consider that input experiment has been generated thanks to the MorrisExperiment and output is evaluated outside the platform. With second constructor, we consider that initial experiment is a regular grid defined in $[0,1]^d$. With thirds constructor, we consider that initial distribution model is uniform with bounds given by the interval argument. Also, the initial experiment is of type regular. With fourth constructor, we fix the initial design which could be an LHS, an optimal LHS defined in $[0,1]^d$. With fifth constructor, we fix the initial design which could be an LHS, an optimal LHS defined in $[0,1]^d$.

Evaluating the Morris test function is very slow

The following code uses vectorized statements and is much faster:

def Morris_function(x):
    delta=np.zeros((20,20,20)); delta[:5,:5,:5].fill(-10)
    	gamma=np.zeros((20,20,20,20)); gamma[:4,:4,:4,:4].fill(5)
    W=np.sum(alpha.reshape(-1,1)*w(x))
    W1=np.sum(beta*(w(x).reshape(1,-1)*w(x))*(np.arange(20).reshape(-1,1)<np.arange(20)))
    W2=np.sum(delta*(w(x).reshape(-1,1,1)*w(x).reshape(1,-1,1)*w(x).reshape(1,1,-1))*(np.arange(20).reshape(-1,1,1)<np.arange(20).reshape(1,-1,1))*(np.arange(20).reshape(1,-1,1)<np.arange(20).reshape(1,1,-1)))
    W3=np.sum(gamma*(w(x).reshape(-1,1,1,1)*w(x).reshape(1,-1,1,1)*w(x).reshape(1,1,-1,1)*w(x).reshape(1,1,1,-1))*(np.arange(20).reshape(-1,1,1,1)<np.arange(20).reshape(1,-1,1,1))*(np.arange(20).reshape(1,-1,1,1)<np.arange(20).reshape(1,1,-1,1))*(np.arange(20).reshape(1,1,-1,1)<np.arange(20).reshape(1,1,1,-1)))
    y=W+W1+W2+W3+float(b0)
    return y

However, the current implementation has no unit tests. This makes changing the implementation unconformable.

version 0.10 available on pip for windows

Hello,

On Pypi for windows distribution, it seems that the latest version of otmorris on PyPi is 0.10.

Mathieu

PlotEE generates an error

The script

import openturns as ot
import otmorris

# Define model
ot.RandomGenerator.SetSeed(1)
g = ot.Function(otmorris.MorrisFunction())

# Number of trajectories
r = 10
# Define experiments in [0,1]^20
# p-levels
p = 5
morris_experiment = otmorris.MorrisExperimentGrid([p] * 20, r)
grid_bound = morris_experiment.getBounds()
X = morris_experiment.generate()

n = X.getSize()

Y = g(X)

# Evaluate Elementary effects (ee)
ee = otmorris.Morris(X, Y, grid_bound)

otmorris.plot_sensitivity.PlotEE(ee)

generates the error:

No handles with labels found to put in legend.

This is otmorris '0.5'

fullDesignSize overflow for large dimension

It can happen fullDesignSize gets overflowed when dealing with a large dimension problem
I propose the following simple fix

void MorrisExperimentGrid::setJumpStep(const Indices & jumpStep)
   // Update the jump step and check that we still might generate N_ trajectories
   // Compute step & number of total possibilities
   // Depending on direction
-  UnsignedInteger fullDesignSize = 2;
+  Scalar fullDesignSize = 2.;
   for (UnsignedInteger k = 0; k < jumpStep.getSize(); ++k)
   {
     const UnsignedInteger one = 1;

User defined input description in plotEE

Some users want to specify the description of the input variables and it is cumbersome to do with the current implementation.

MorrisFunction undocumented

The MorrisFunction function is used in the examples. It has docstrings, but it does not appear in the Sphynx documentation.

otmorris depends on amslatex which is not checked at install time

See https://openturns.discourse.group/t/error-latex-with-plotee-morris-method/165 for the details. IMO it is crazy to add a dependence to amstex only to get two greek letters in boldface!

Installation on Google colab

Some users want to use otmorris from Google colab:

https://colab.research.google.com

But colab does not know about conda. Is there any way to get around this?

Module installation

Installing the otmorris module is always a nightmare, because there is only a conda package.

Installing it requires to install the openturns version from conda, which is very long.
To make the plots work, another module is required, which is not automatically installed:

sudo apt-get install dvipng

Even with this setup, I have the following error that I do not understand:

$ conda install -c conda-forge otmorris
Collecting package metadata (current_repodata.json): done
Solving environment: done
==> WARNING: A newer version of conda exists. <==
  current version: 4.8.3
  latest version: 4.9.2
Please update conda by running
    $ conda update -n base -c defaults conda
# All requested packages already installed.
$ python
Python 3.8.3 (default, May 19 2020, 18:47:26) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import otmorris
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ModuleNotFoundError: No module named 'otmorris'

My questions:

What is the cause of that error?
Is there any way to create a pip package? It would be much easier and faster.

MorrisExperimentGrid hangs

Seems the while loop hangs when there are not enough trajectories under certain conditions (dim=1, levels=[5], n_traj=10):

from __future__ import print_function
import openturns as ot
import otmorris

# Use Case 1 : define a k-grid level (so delta = 1/(k-1))
r = 10
morris_experiment = otmorris.MorrisExperimentGrid([5], r)
X = morris_experiment.generate()
print("Morris experiment generated from grid = ", X)

Here we enforce unique trajectories.
solution: compute the number of possible trajectories: (d-1)^d modulo the jumpstep wrt out of bound.

The generate method of MorrisExperimentGrid fails

The following script fails with OT 1.14:

import openturns as ot
import otmorris

# Define model
ot.RandomGenerator.SetSeed(1)
b0 = ot.DistFunc.rNormal()
alpha = ot.DistFunc.rNormal(10)
beta = ot.DistFunc.rNormal(6*14)
gamma = ot.DistFunc.rNormal(20*14)
g = ot.Function(otmorris.MorrisFunction(alpha, beta, gamma, b0))

dim = g.getInputDimension()

# Number of trajectories
r = 10
# Define experiments in [0,1]^20
# p-levels
p = 5
morris_experiment = otmorris.MorrisExperimentGrid([p] * dim, r)
grid_bound = morris_experiment.getBounds()
X = morris_experiment.generate()
print(X)

with

double free or corruption (out)
Aborted (core dumped)

This test is with

openturns : 1.14
otmorris : 0.7

MorrisExperimentLHS segfaults

The following script segfaults with otmorris 0.7 on OS X installed via conda:

import openturns as ot
import otmorris
X = ot.LHSExperiment(ot.ComposedDistribution([ot.Uniform(0,1)]*4), 50, False, False).generate()
morris_experiment = otmorris.MorrisExperimentLHS(X, 10)
morris_experiment.generate()

The same code using cd0423c has no issue on OS X (compiled on by laptop).

It seems that bug is conda related?