Customizing your own model parameters

This page delves into some of the more powerful things you can do with pipelines. It is aimed at people who have already completed the How to write advanced pipelines tutorial.

Pipeline model parameters

You should be aware that each pipeline step has parameters, but the pipeline model itself can also have its own parameters. Unlike step parameters, you have complete control over defining the model parameters yourself.

The model parameters are essentially custom variables that get replaced at run-time. These variables are denoted by a $ in the pipeline file.

The model parameters can customize the behaviour of the model across all input data. Model parameters can encapsulate an assumption you are making about your model, which you may want to vary.

As a simple example, say we only wanted to run our model for one particular construction type at a time. But we wanted to be able to vary that construction type easily, without editing the pipeline file each time. We can turn the construction type we filter on into a construction parameter, e.g.

input('assets.csv', name: 'asset') -> filter(filter: asset.construction = $construction)

When we run our model, we can then specify what specific construction type we are interested in, e.g.

riskscape model run my_pipeline --param "construction='steel'"

You can define the default values for your parameters in your model’s INI definition. Just use param.<name> = <default>, e.g.

[model my_pipeline]
framework = pipeline
location = my_pipeline.txt
param.construction = 'timber'

Function parameters

Model parameters also work if the assumption you are making is in your function. You simply need to pass the model parameter through to your function.

To do this you add an extra argument to your function. The new argument will be a Struct and contain the attributes you want to vary.

For example, say our Kaijū function was making an assumption on how resilient timber buildings are to Kaijū attacks. We can modify our function to take an additional ‘options’ argument, which is a Struct with a timber_resilience attribute. The (abridged) function would look like this:

from nz.org.riskscape.engine.types import Types
from nz.org.riskscape.engine.types import Struct

ID = 'kaiju_stomp'
DESCRIPTION = 'Models damage from a Kaiju stomping a building'

ARGUMENT_TYPES = ['building', 'kaiju_attack', \
                  Struct.of('timber_resilience', Types.INTEGER) ]

RETURN_TYPE = 'building_attack_outcome'

def function(building, stomp, options):
    if building.get('construction') == 'timber':
        # the resilience for timber buildings can be passed
        # into our function, making it easier to vary it
        resilience = options.get('timber_resilience')
    elif building.get('construction') == 'concrete':
        # whereas the resilience for concrete buildings is
        # still hard-coded in the function itself
        resilience = 5
    # ...

Then to call our function from our pipeline code, we create a new Struct with a timber_resilience attribute. In this case, the value for timber_resilience is a model parameter that we can now change on the fly whenever we run our model.

select({*,
        kaiju_stomp(asset, hazard, { timber_resilience: $timber_resilience }) as damage
      ) as compute consequence

The ‘options’ Struct could hold many different attributes, if there are many assumptions in your function that you want to vary.

A working example

Click Here to download a working example of a parameterized pipeline.

Open the pipeline.txt and project.ini files and familiarize yourself with them. You can see that the model defines ‘resilience’ parameters for the three construction materials of interest. These parameters then get passed directly to the kaiju_stomp function.

To run the model with default parameter values, use the command:

riskscape model run demo

Reproducible models

Model parameters can make it harder to tell later what actual parameter values were used to produce a certain set of results. Fortunately, RiskScape always saves the actual parameter values it used in the output directory.

Have a look at the output directory created by the last ‘model run’ command. It should contain a pipeline.txt file - this contains the pipeline code with the $ parameters replaced with the actual values used by the ‘model run’ command.

Any piece of raw pipeline code (i.e. without parameters) can be executed using the riskscape pipeline eval command. So you can use the pipeline.txt file in the output directory to re-run the exact same pipeline again in the future, e.g.

riskscape pipeline eval <pipeline.txt>

Try this now and check you get the same results.

Parameter INI files

When your model has several different parameters, rather than specifying each parameter on the command line, you can just specify one INI file that contains all the parameter values you want to use.

To pass the INI file to the ‘model run’ command, use the --parameters CLI option (note the ‘s’ on the end). For example:

riskscape model run demo --parameters stronger-resilience.ini

Using INI files can be helpful if you want to vary several different assumptions in a related way. For example, this project contains two different INI files that make different assumptions about a building’s resilience: stronger-resilience.ini and weaker-resilience.ini. Try running the model with each of these INI files and see what difference it makes to the overall damage.