Getting started


pyMOR can be easily installed via pip:

pip install --upgrade pip  # make sure that pip is reasonably new
pip install pymor[full]

For Linux we provide binary wheels, so no further system packages should be required. The following optional packages must be installed separately:

# for support of MPI distributed models and parallelization of
# greedy algorithms (requires MPI development headers and a C compiler)
pip install mpi4py

# dense matrix equation solver for system-theoretic MOR methods,
# required for H-infinity norm calculation (requires OpenBLAS headers and a Fortran compiler)
pip install slycot

# sparse matrix equation solver for system-theoretic MOR methods
# (other backends available)
# download and install pymess wheel for your architecture

We recommend installation of pyMOR in a virtual environment. Please take a look at our README file for more detailed installation instructions and a guide to setup a development environment for working on pyMOR itself.

Trying it out

While we consider pyMOR mainly as a library for building MOR applications, we ship a few example scripts. These can be found in the src/pymordemos directory of the source repository (some are available as Jupyter notebooks in the notebooks directory). Try launching one of them using the pymor-demo script:

pymor-demo thermalblock --plot-err --plot-solutions 3 2 3 32

The demo scripts can also be launched directly from the source tree:

./ --plot-err --plot-solutions 3 2 3 32

This will reduce the so called thermal block problem using the reduced basis method with a greedy basis generation algorithm. The thermal block problem consists in solving the stationary heat equation

- ∇ ⋅ [ d(x, μ) ∇ u(x, μ) ] = 1     for x in Ω
                  u(x, μ)   = 0     for x in ∂Ω

on the domain Ω = [0,1]^2 for the unknown u. The domain is partitioned into XBLOCKS x YBLOCKS blocks (XBLOCKS and YBLOCKS are the first two arguments to The thermal conductivity d(x, μ) is constant on each block (i,j) with value μ_ij:

|        |        |        |
|  μ_11  |  μ_12  |  μ_13  |
|        |        |        |
|        |        |        |
|  μ_21  |  μ_22  |  μ_23  |
|        |        |        |

The real numbers μ_ij form the XBLOCKS x YBLOCKS - dimensional parameter on which the solution depends.

Running will first produce plots of two detailed solutions of the problem for different randomly chosen parameters using linear finite elements. (The size of the grid can be controlled via the --grid parameter. The randomly chosen parameters will actually be the same for each run, since a the random generator is initialized with a fixed default seed in new_random_state.)

After closing the window, the reduced basis for model order reduction is generated using a greedy search algorithm with error estimator. The third parameter SNAPSHOTS of determines how many different values per parameter component μ_ij should be considered. I.e. the parameter training set for basis generation will have the size SNAPSHOTS^(XBLOCKS x YBLOCKS). After the basis of size 32 (the last parameter) has been computed, the quality of the obtained reduced model (on the 32-dimensional reduced basis space) is evaluated by comparing the solutions of the reduced and detailed models for new, randomly chosen parameters. Finally, plots of the detailed and reduced solutions, as well as the difference between the two, are displayed for the random parameter which maximises reduction error.

The thermalblock demo explained

In the following we will walk through the thermal block demo step by step in an interactive Python shell. We assume that you are familiar with the reduced basis method and that you know the basics of Python programming as well as working with NumPy. (Note that our code will differ a bit from as we will hardcode the various options the script offers and leave out some features.)

First, start a Python shell. We recommend using IPython


You can paste the following input lines starting with >>> by copying them to the system clipboard and then executing


inside the IPython shell.

First, we will import the most commonly used methods and classes of pyMOR by executing:

>>> from pymor.basic import *

Next we will instantiate a class describing the analytical problem we want so solve. In this case, a thermal_block_problem:

>>> p = thermal_block_problem(num_blocks=(3, 2))

We want to discretize this problem using the finite element method. We could do this by hand, creating a Grid, instatiating DiffusionOperatorP1 finite element diffusion operators for each subblock of the domain, forming a LincombOperator to represent the affine decomposition, instantiating a L2ProductFunctionalP1 as right hand side, and putting it all together into a StationaryDiscretization. However, since thermal_block_problem returns a StationaryProblem, we can use a predifined discretizer to do the work for us. In this case, we use discretize_stationary_cg:

>>> d, d_data = discretize_stationary_cg(p, diameter=1./100.)

d is the StationaryDiscretization which has been created for us, whereas d_data contains some additional data, in particular the Grid and the BoundaryInfo which have been created during discretization. We can have a look at the grid,

>>> print(d_data['grid'])
Tria-Grid on domain [0,1] x [0,1]
x0-intervals: 100, x1-intervals: 100
elements: 40000, edges: 60200, vertices: 20201

and, as always, we can display its class documentation using help(d_data['grid']).

Let’s solve the thermal block problem and visualize the solution:

>>> U = d.solve([1.0, 0.1, 0.3, 0.1, 0.2, 1.0])
>>> d.visualize(U, title='Solution')
01:11 StationaryDiscretization: Solving ThermalBlock((3, 2))_CG for {diffusion: [1.0, 0.1, 0.3, 0.1, 0.2, 1.0]} ...

Each class in pyMOR that describes a Parameter dependent mathematical object, like the StationaryDiscretization in our case, derives from Parametric and determines the Parameters it expects during __init__ by calling build_parameter_type. The resulting ParameterType is stored in the object’s parameter_type attribute. Let us have a look:

>>> print(d.parameter_type)
{diffusion: (2, 3)}

This tells us, that the Parameter which solve expects should be a dictionary with one key 'diffusion' whose value is a NumPy array of shape (2, 3), corresponding to the block structure of the problem. However, by using the parse_parameter method, pyMOR is smart enough to correctly parse the input [1.0, 0.1, 0.3, 0.1, 0.2, 1.0].

Next we want to use the greedy algorithm to reduce the problem. For this we need to choose a reductor which will keep track of the reduced basis and perform the actual RB-projection. We will use CoerciveRBReductor, which will also assemble an error estimator to estimate the reduction error. This will significantly speed up the basis generation, as we will only need to solve the high-dimensional problem for those parameters in the training set which are actually selected for basis extension. To control the condition of the reduced system matrix, we must ensure that the generated basis is orthonormal w.r.t. the H1_0-product on the solution space. For this we pass the h1_0_semi_product attribute of the discretization as inner product to the reductor, which will also use it for computing the Riesz representatives required for error estimation. Moreover, we have to provide the reductor with a ParameterFunctional which computes a lower bound for the coercivity of the problem for a given parameter.

>>> reductor = CoerciveRBReductor(
...     d,
...     product=d.h1_0_semi_product,
...     coercivity_estimator=ExpressionParameterFunctional('min(diffusion)', d.parameter_type)
... )

Moreover, we need to select a Parameter training set. The discretization d already comes with a ParameterSpace which it has inherited from the analytical problem. We can sample our parameters from this space, which is a CubicParameterSpace. E.g.:

>>> samples = d.parameter_space.sample_uniformly(4)
>>> print(samples[0])
{diffusion: [0.1, 0.1, 0.1, 0.1, 0.1, 0.1]}

Now we start the basis generation:

>>> greedy_data = greedy(d, reductor, samples,
...                      use_estimator=True,
...                      max_extensions=32)
16:52 greedy: Started greedy search on 4096 samples
16:52 greedy: Reducing ...
16:52 |   CoerciveRBReductor: RB projection ...
16:52 |   CoerciveRBReductor: Assembling error estimator ...
16:52 |   |   ResidualReductor: Estimating residual range ...
16:52 |   |   |   estimate_image_hierarchical: Estimating image for basis vector -1 ...
16:52 |   |   |   estimate_image_hierarchical: Orthonormalizing ...
16:52 |   |   ResidualReductor: Projecting residual operator ...
16:52 greedy: Estimating errors ...
16:55 greedy: Maximum error after 0 extensions: 1.8745731821515579 (mu = {diffusion: [0.1, 0.1, 0.1, 0.1, 0.1, 0.1]})
16:55 greedy: Computing solution snapshot for mu = {diffusion: [0.1, 0.1, 0.1, 0.1, 0.1, 0.1]} ...
16:55 |   StationaryDiscretization: Solving ThermalBlock((3, 2))_CG for {diffusion: [0.1, 0.1, 0.1, 0.1, 0.1, 0.1]} ...
16:55 greedy: Extending basis with solution snapshot ...
18:57 greedy: Maximum number of 32 extensions reached.
18:57 greedy: Reducing once more ...
18:57 |   CoerciveRBReductor: RB projection ...
18:57 |   CoerciveRBReductor: Assembling error estimator ...
18:57 |   |   ResidualReductor: Estimating residual range ...
18:57 |   |   |   estimate_image_hierarchical: Estimating image for basis vector 31 ...
18:57 |   |   |   estimate_image_hierarchical: Orthonormalizing ...
18:57 |   |   |   |   gram_schmidt: Removing vector 180 of norm 1.7588304501544013e-15
18:57 |   |   |   |   gram_schmidt: Orthonormalizing vector 181 again
18:57 |   |   |   |   gram_schmidt: Orthonormalizing vector 182 again
18:57 |   |   |   |   gram_schmidt: Orthonormalizing vector 183 again
18:58 |   |   |   |   gram_schmidt: Orthonormalizing vector 184 again
18:58 |   |   |   |   gram_schmidt: Orthonormalizing vector 185 again
18:58 |   |   |   |   gram_schmidt: Orthonormalizing vector 186 again
18:58 |   |   ResidualReductor: Projecting residual operator ...
18:58 greedy: Greedy search took 126.14163041114807 seconds

The max_extensions parameter defines how many basis vectors we want to obtain. greedy_data is a dictionary containing various data that has been generated during the run of the algorithm:

>>> print(greedy_data.keys())
dict_keys(['rd', 'max_errs', 'extensions', 'max_err_mus', 'time'])

The most important items is 'rd' which holds the reduced Discretization obtained from applying our reductor with the final reduced basis.

>>> rd = greedy_data['rd']

All vectors in pyMOR are stored in so called VectorArrays. For example the solution U computed above is given as a VectorArray of length 1. For the reduced basis we have:

>>> print(type(reductor.RB))
<class 'pymor.vectorarrays.numpy.NumpyVectorArray'>
>>> print(len(reductor.RB))
>>> print(reductor.RB.dim)

Let us check if the reduced basis really is orthonormal with respect to the H1-product. For this we use the apply2 method:

>>> import numpy as np
>>> gram_matrix = reductor.RB.gramian(d.h1_0_semi_product)
>>> print(np.max(np.abs(gram_matrix - np.eye(32))))

Looks good! We can now solve the reduced model for the same parameter as above. The result is a vector of coefficients w.r.t. the reduced basis, which is currently stored in rb. To form the linear combination, we can use the reconstruct method of the reductor:

>>> u = rd.solve([1.0, 0.1, 0.3, 0.1, 0.2, 1.0])
>>> print(u)
[[  5.79477471e-01   5.91289054e-02   1.89924036e-01   1.89149529e-02
    1.81103127e-01   2.69920752e-02  -1.79611519e-01   7.99676272e-03
    1.54092560e-01   5.76326362e-02   1.97982347e-01  -2.13775254e-02
    3.12892660e-02  -1.27037440e-01  -1.51352508e-02   3.36101087e-02
    2.05779889e-02  -4.96445984e-03   3.21176662e-02  -2.52674851e-02
    2.92150040e-02   3.23570362e-03  -4.14288199e-03   5.48325425e-03
    4.10728945e-03   1.59251955e-03  -9.23470903e-03  -2.57483574e-03
   -2.52451212e-03  -5.08125873e-04   2.71427033e-03   5.83210112e-05]]
>>> U_red = reductor.reconstruct(u)
>>> print(U_red.dim)

Finally we compute the reduction error and display the reduced solution along with the detailed solution and the error:

>>> ERR = U - U_red
>>> print(ERR.norm(d.h1_0_semi_product))
>>> d.visualize((U, U_red, ERR),
...             legend=('Detailed', 'Reduced', 'Error'),
...             separate_colorbars=True)

We can nicely observe that, as expected, the error is maximized along the jumps of the diffusion coefficient.

Learning more

As a next step, you should read our Technical Overview which discusses the most important concepts and design decisions behind pyMOR. After that you should be ready to delve into the reference documentation.

Should you have any problems regarding pyMOR, questions or feature requests, do not hesitate to contact us at our mailing list!