XML workflow for Python plugin with external measured BRDF files

[Daavid739]

Hello all,

I'm trying to find the correct workflow for using a custom data-driven BSDF, written in Python, within an XML scene file. Specifically, my goal is to use my own measured BRDF data, which I have in .h5 or .npy file formats, to drive a custom BSDF plugin. The documentation is very clear on how to instantiate plugins within Python using mi.load_dict() and hard-coded parameters, like this example:

my_bsdf = mi.load_dict({ 'type' : 'mybsdf', 'tint' : [0.2, 0.9, 0.2] }) 

However, the documentation doesn't seem to explicitly cover the workflow for two related use cases:
1. How to instantiate that same custom Python plugin from an XML file.
2. How to pass an external data file path (like my .npy or .h5 files) as a parameter from the XML, instead of just hard-coded values.
Based on my experiments, I believe the process is as follows. Could you please confirm if this is the correct and intended approach?

A Python class that is designed to accept a filename from the props object and load the data using NumPy or h5py.

class MyMeasuredBSDF(mi.BSDF):
    def __init__(self, props):
        mi.BSDF.__init__(self, props)
        filename = props['my_data_file']
        self.data = np.load(filename)

A main Python script that registers the class before loading the XML scene.

import mitsuba as mi
from my_plugin import MyMeasuredBSDF
mi.set_variant('scalar_rgb')
mi.register_bsdf("my_material", MyMeasuredBSDF)
scene = mi.load_file("scene.xml") # <-- Load XML after registration

The XML uses the registered name as the type and passes the file path using a tag.

<scene version="3.0.0">
    <shape type="sphere">
        <bsdf type="my_material">
            <string name="my_data_file" value="path/to/data.npy"/>
        </bsdf>
    </shape>
</scene>

My main question is: Is this three-step process the correct way to link a custom, data-driven Python plugin (using formats like .npy or .h5) to an XML scene?

@wjakob @nroussel

[merlinND]

Hello @Daavid739,

Yes, the process you describe sounds right:

1. Make your custom Python plugin read the filename as a string from props
2. Make sure that your plugin is registered before calling load_file/load_dict
3. In the XML file, set type="" to your plugin name and nest a string property with the correct name to pass your filename.