Issue running Complete Example

[LSgeo]

Trying to get started with the project, running the Complete Example at https://opensource.michelin.io/TorchSOM/getting_started/quickstart.html#complete-example. The SOM trains, but has an error in a visualisation step. The indices_tensor is floating point values, so it doesn't seem appropriate to cast to an integer type.

Expected behaviour: Complete example in documentation runs without error.

Training SOM: 100%|█████████████████████████████████████████| 100/100 [00:13<00:00,  7.65epoch/s]
Traceback (most recent call last):
  File "...\som\som_poc.py", line 48, in <module>
    viz.plot_all(
  File "...\.virtualenvs\som\Lib\site-packages\torchsom\visualization\base.py", line 193, in plot_all
    self._visualizer.plot_metric_map(
  File "...\.virtualenvs\som\Lib\site-packages\torchsom\visualization\base_visualizer.py", line 211, in plot_metric_map
    metric_map = self.som.build_map(
                 ^^^^^^^^^^^^^^^^^^^
  File "...\.virtualenvs\som\Lib\site-packages\torchsom\core\som.py", line 642, in build_map
    return map_function(self, bmus_data_map, target, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "...\.virtualenvs\som\Lib\site-packages\torchsom\utils\maps.py", line 241, in build_metric_map
    target_values = target[indices_tensor].float()
                    ~~~~~~^^^^^^^^^^^^^^^^
IndexError: tensors used as indices must be long, int, byte or bool tensors

Python environment, torch/cuda version 2.13.0+cu130

pip list
Package           Version
----------------- ------------
annotated-types   0.8.0
colorama          0.4.6
contourpy         1.3.3
cycler            0.12.1
faiss-cpu         1.14.3
filelock          3.32.0
fonttools         4.63.0
fsspec            2026.7.0
hdbscan           0.8.44
Jinja2            3.1.6
joblib            1.5.3
kiwisolver        1.5.0
MarkupSafe        3.0.3
matplotlib        3.8.4
mpmath            1.3.0
narwhals          2.24.0
networkx          3.6.1
numpy             2.5.1
packaging         26.2
pillow            12.3.0
pip               24.3.1
pydantic          2.13.4
pydantic_core     2.46.4
pyparsing         3.3.2
python-dateutil   2.9.0.post0
scikit-learn      1.9.0
scipy             1.18.0
setuptools        83.0.0
six               1.17.0
sympy             1.14.0
threadpoolctl     3.6.0
torch             2.13.0+cu130
torchsom          1.3.0
tqdm              4.70.0
typing_extensions 4.16.0
typing-inspection 0.4.2

[LSgeo]

As a newcomer to the project I have troubleshooted with the help of ChatGPT and diagnosed that is is simply a missing return_indices=True kwarg in the example script provided in the rst files.

bmus_map = som.build_map("bmus_data", data=data) defaults to have return_indices=False, so bmus_map contains the actual feature vectors rather than sample indices.

The example (in https://github.com/michelin/TorchSOM/blob/main/docs/source/getting_started/quickstart.rst#complete-example) should instead use:

bmus_map = som.build_map(
    "bmus_data",
    data=data,
    return_indices=True,
)

I confirmed that this produces integer sample indices and resolves the error.

I've seen other projects use python doctest to catch these issues - it might be a useful addition to ponder / first issue for someone.