Consultation about "trainable_materials.py"

[kevcfGit]

Good day!

I have been trying to implement the material training approach presented in "Learning Radio Environments by Differentiable Ray Tracing" paper but not basing my experiment in Compute_Paths() but instead on Coverage_map() in order to see how well calibration behaves; I have tried to adapt "trainable_materials.py" into Coverage_map() function but i keep getting erros.

Since I'm working with coverage_map() rather than compute_paths(), I understand that ProxyRadioMaterial is not automatically used during ray-object interactions. To resolve this, I followed a hint in your documentation and defined:

scene.radio_material_callable = my_callable

The callable follows the required (object_ids, points) → (eta, scatter, xpd) signature and uses tf.numpy_function internally to map global object_ids to embedding indices via a Python dict (obj_id_to_idx).

• Problem:
  Although I do see TrainableMaterials.call() being triggered, the execution crashes deep inside solver_cm.py, usually at a scattering or diffraction-related step. The traceback points to tf.numpy_function and reports:

  1. TypeError: unhashable type: 'numpy.ndarray'
  2. Or: only length-1 arrays can be converted to a Python scalar
  3. Or: Outer dimensions of indices and update must match during tensor_scatter_nd_update

• I’ve tried:

  1. Using int(i) or i.item() inside the lambda used in tf.numpy_function
  2. Converting ids to .tolist() before indexing
  3. Reshaping tensors before tensor_scatter_nd_update
  4. Wrapping everything with @tf.function and testing in both eager and graph mode

• My questions:

  1. Is there an officially recommended way to integrate custom trainable material behavior with coverage_map()? (similar to the one presented in "Learning Radio Environments by Differentiable Ray Tracing" paper.
  2. Is scene.radio_material_callable expected to be used instead of ProxyRadioMaterial, or in addition to?
  3. Are there any internal tools/helpers to map object IDs during graph execution (e.g., tf.lookup)? Or should we modify the source to inject local material indices during ray tracing?

I was wondering if you could point the right direction for me to follow in this implementation or state that this is not suitable for coverage_ma().
(I have already reviewed instant radio maps calibration notebook but for personal purposes, i need the features provided in SIONNA)
Validation_SyntheticTruth.txt

Important to mention that i am currently using: Sionna 0.19.2

Approach:

Source code:
https://github.com/NVlabs/diff-rt-calibration/blob/main/code/trainable_materials.py