File Passing Tutorial - Passing References to Files Between Components

References to files can be passed between Components in OpenMDAO using variables called FileRefs. A FileRef is just an object that contains a file name and an absolute file path which is calculated by the framework. The fname attribute can be a simple name or a name that includes a relative or absolute directory path.

Calculation of Absolute Directory Paths

During setup, OpenMDAO determines the absolute file system path for each FileRef variable based on the directory path of the Component that contains it. The process works like this:

  1. Starting at the root System of the tree, we calculate its absolute directory based on its ‘directory’ attribute. If the ‘directory’ attribute is empty, then the absolute directory is just the current working directory. If directory contains a relative pathname, then the absolute directory is the current working directory plus the relative path. If directory is already an absolute path, then we just use that.
  2. For each child System in the tree, we calculate its absolute directory based on the absolute directory we’ve already calculated for its parent System, in the same manner as in step 1, except that instead of the current working directory, we use the parent absolute directory as our starting point.
  3. In each Component we encounter as we traverse the tree, after we’ve calculated its absolute directory, we look for any FileRef variables it may have. We set the absolute directory for each FileRef in a similar way as in step 2, but in this case the Component is the parent, so we use its absolute directory as our starting point in determining the absolute path of the FileRef.

Note

Sometimes you may not want to hard-code the directory name. If you want to delay picking a name until runtime, you can specify directory as a function. If directory is a function, we will call that function, passing in the rank for the current process. That function should return a string containing either a relative or absolute path, which we will resolve to an absolute directory as mentioned above.

Using FileRefs

So lets make some components that pass FileRefs between them. We’ll just use ascii files here to keep things as simple as possible, but FileRefs can be binary if you set binary=True in the metadata when you add them to a component.

First, we’ll make a simple component that takes a single parameter, does a simple calculation, then writes the result to a file.

from openmdao.api import Problem, Group, Component, FileRef

class FoutComp(Component):
    """A component that writes out a file containing a number."""

    def __init__(self):
        super(FoutComp, self).__init__()

        # add a simple parameter that we can use to calculate the
        # number we write to our output file
        self.add_param('x', 1.0)

        # add an output FileRef for our output file 'dat.out'
        self.add_output("outfile", FileRef("dat.out"))

    def solve_nonlinear(self, params, unknowns, resids):
        # do some simple calculation
        val = params['x'] * 2.0 + 1.0

        # write the new value to our output FileRef
        with unknowns['outfile'].open('w') as f:
            f.write(str(val))

Now we need a component to read a number from our first component’s output file and use that to calculate a new number.

class FinComp(Component):
    """A component that reads a file containing a number."""

    def __init__(self):
        super(FinComp, self).__init__()

        # add an input FileRef for our input file 'dat.in'
        self.add_param("infile", FileRef("dat.in"))

        # here's the output we'll calculate using the number we read
        # from our input FileRef
        self.add_output('y', 1.0)

    def solve_nonlinear(self, params, unknowns, resids):
        # read the number from our input FileRef
        with params['infile'].open('r') as f:
            val = float(f.read())

        # now calculate our new output value
        unknowns['y'] = val + 7.0

Now we have our two file transferring components, so we can build our model.

p = Problem(root=Group())
outfilecomp = p.root.add("outfilecomp", FoutComp())
infilecomp = p.root.add("infilecomp", FinComp())

# connect our two FileRefs together
p.root.connect("outfilecomp.outfile", "infilecomp.infile")

p.setup()

We’ll set a value of 3.0 in our first component’s x value. That should give us a y value in our second component of 14.0.

p['outfilecomp.x'] = 3.0

p.run()

print(p['infilecomp.y'])
14.0

In this example, our files were both in the same directory, but you can control where they are found by modifying the directory attribute of systems in the tree. For example, if we wanted outfilecomp.outfile to be located in a sub1 subdirectory, we could do the following:

p = Problem(root=Group())
outfilecomp = p.root.add("outfilecomp", FoutComp())

# specify the subdirectory here
outfilecomp.directory = 'sub1'

# since 'sub1' doesn't exist, we need to tell the component to create it.
# otherwise, we'll get an error that the directory doesn't exist.
outfilecomp.create_dirs = True

infilecomp = p.root.add("infilecomp", FinComp())

# connect our two FileRefs together
p.root.connect("outfilecomp.outfile", "infilecomp.infile")

p.setup()

Notice that none of the code in our components or any of our other configuration code has changed. When we run this problem, we get the same answer as before.

p['outfilecomp.x'] = 3.0

p.run()

print(p['infilecomp.y'])
14.0

FileRefs under MPI

When running under MPI, there are certain situations where you may need to create subdirectories dynamically based on the rank of the current MPI process. You can accomplish that by assigning a function to a system’s directory instead of just a simple string. For example, suppose we had a group in our model that we wanted to perform parallel finite difference on, and that group happened to have output FileRefs in it. In that situation, different MPI processes would try to write to the same output file at the same time. In order to prevent this, we can specify that in each MPI process, our group will have a directory specific to that process. Assigning directory to a function instead of a string will let us do that. For example, let’s say we want our group to write its files in a subdirectory called ‘foo_n’, where ‘n’ is the rank of the current process. In that case, setting our group’s directory would look like this:

mygrp.directory = lambda rank: "foo_%d" % rank
mygrp.create_dirs = True  # create the directories if they don't exist

The function you assign to directory should expect a single argument that is the rank of the current process, and it should return the desired directory string. Note that it’s also valid to assign a method of your component to directory if you happen to need more information than just the rank in order to determine the directory name. For example:

class MyComp(FoutComp):
    def get_dirname(self, rank):
        return "%s_%d" % (self.name, rank)

mycomp = MyComp()
mycomp.directory = mycomp.get_dirname
mycomp.create_dirs = True

Tags

Tutorials