Automatic creation of terratorch iterate config from a terratorch config

[paolofraccaro]

One of the barriers of using terratorch iterate is the creation of a terratorch iterate config from a terratorch config. We should have an automatic script/function that does that.

[leotizzei]

Hi @paolofraccaro,

I moved the script that you implemented to this directory. Then I refactored it for the following reasons :

• fix some issues (e.g., missing fields)
• integrate with click, a python CLI lib, which allows users to specify directory, template and output via terminal
• improve documentation

I documented how to run this script here.

Caveats:

• the refactored script generates the config file, but I have not tested this config file yet, because I don't have access to the datasets. But I tested against the files generated by your original script and they are very similar
• it does not work with the terratorch config file located in terratorch repoterratorch/tests/resources/configs/. There is still room to make this script more generic, that is, so it works with different config files. It seems that current implementation works well for geobench_v2 examples.

[leotizzei]

this is the error message that I'm getting when I run tt-iterate using this config. I think we need to change the way this script sets task type

[I 2025-08-06 15:44:31,039] A new study created in RDB with name: manufactured-finetune
INFO: Seed set to 42
[W 2025-08-06 15:44:32,190] Trial 0 failed with parameters: {'lr': 1.7971004101534673e-06} because of the following error: Exception('num_classes must be defined for segmentation task').
Traceback (most recent call last):
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/optuna/study/_optimize.py", line 197, in _run_trial
    value_or_values = func(trial)
                      ^^^^^^^^^^^
  File "/u/ltizzei/Projects/Orgs/IBM/terratorch-iterate/benchmark/model_fitting.py", line 446, in fit_model_with_hparams
    return fit_model(
           ^^^^^^^^^^
  File "/u/ltizzei/Projects/Orgs/IBM/terratorch-iterate/benchmark/model_fitting.py", line 357, in fit_model
    lightning_task = lightning_task_class(**task.terratorch_task)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/tasks/segmentation_tasks.py", line 143, in __init__
    super().__init__(
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/tasks/base_task.py", line 38, in __init__
    super().__init__()
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/torchgeo/trainers/base.py", line 42, in __init__
    self.configure_models()
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/tasks/base_task.py", line 51, in configure_models
    self.model: Model = self.model_factory.build_model(
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/models/prithvi_model_factory.py", line 96, in build_model
    return self._factory.build_model(task,
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/models/encoder_decoder_factory.py", line 196, in build_model
    return _build_appropriate_model(
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/models/encoder_decoder_factory.py", line 254, in _build_appropriate_model
    return PixelWiseModel(
           ^^^^^^^^^^^^^^^
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/models/pixel_wise_model.py", line 58, in __init__
    self._get_head(task, decoder.out_channels, head_kwargs) if not decoder_includes_head else nn.Identity()
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/u/ltizzei/.pyenv/versions/3.12.8/envs/it/lib/python3.12/site-packages/terratorch/models/pixel_wise_model.py", line 154, in _get_head
    raise Exception(msg)
Exception: num_classes must be defined for segmentation task
[W 2025-08-06 15:44:32,197] Trial 0 failed with value None.

[leotizzei]

@paolofraccaro , I've copied two config files from terratorch, converted them using the script (this file and this file) and then run terratorch iterate. Both tests passed.

[paolofraccaro]

This is great! I would not call the script build_geobench_configs! Rather build iterate config. We could have a shortcut maybe terratorch iterate --build_iterate_config --config x.yaml?

[leotizzei]

Hi @paolofraccaro , that's a good idea! I'll change main.py module to allow users to run the script using terratorch iterate. Your previous script allows users to convert multiple terratorch's configs into one or more iterate's config. The last script that you shared with me does not allow that. My understanding is that it is still useful to allow users to convert multiple terratorch's configs. If so, then users would run:

 terratorch iterate --build_iterate_config --input <directory-that-contain-terratorch-configs> --ouput <directory-to-store-iterate-configs> --template <path-to-template>

but it would also be possible to convert a single config file:

 terratorch iterate --build_iterate_config --input <terratorch-config-file> --ouput <iterate-config-file> --template <path-to-template>

please let me know your thoughts on this issue