Hi and thanks for releasing TIRAuxCloud and the pretrained models.
While reproducing the metrics, I noticed that the current README assumes users already know how to:
Organize the dataset locally from the Hugging Face structure
Prepare a working saved_model_run.json
Call model_test.py and interpret the outputs.
To make the repo more beginner-friendly and to facilitate integration into more general AI pipelines (e.g. for GSOC-style frameworks for thermal satellite payloads), I’d like to add a small end‑to‑end example under an examples/ folder that:
Downloads a small subset of TIRAuxCloud from Hugging Face (e.g. a few tiles from the Landsat subset).
Provides a minimal example_saved_model_run.json aligned with one of the released pretrained models.
Calls model_test.py programmatically and saves metrics (precision/recall/F1/IoU) to a JSON file.
Optionally plots a simple confusion matrix for the cloud vs non-cloud mask.
This would not change the main evaluation logic, only wrap it in a reproducible, documented script and example config.
If this sounds useful, I’m happy to implement it and open a PR.
Hi and thanks for releasing TIRAuxCloud and the pretrained models.
While reproducing the metrics, I noticed that the current README assumes users already know how to:
Organize the dataset locally from the Hugging Face structure
Prepare a working saved_model_run.json
Call model_test.py and interpret the outputs.
To make the repo more beginner-friendly and to facilitate integration into more general AI pipelines (e.g. for GSOC-style frameworks for thermal satellite payloads), I’d like to add a small end‑to‑end example under an examples/ folder that:
Downloads a small subset of TIRAuxCloud from Hugging Face (e.g. a few tiles from the Landsat subset).
Provides a minimal example_saved_model_run.json aligned with one of the released pretrained models.
Calls model_test.py programmatically and saves metrics (precision/recall/F1/IoU) to a JSON file.
Optionally plots a simple confusion matrix for the cloud vs non-cloud mask.
This would not change the main evaluation logic, only wrap it in a reproducible, documented script and example config.
If this sounds useful, I’m happy to implement it and open a PR.