Skip to content

PR_BODY.md

Latest commit

 

History

History
24 lines (20 loc) · 1.94 KB

PR_BODY.md

File metadata and controls

24 lines (20 loc) · 1.94 KB
Raw

Summary

Implemented a comprehensive pytest suite and resolved bugs in the training metrics to prepare the data and modeling pipeline. We also removed physically-invalid augmentations from the data pipeline and fixed environment build errors.

Problem Addressed

  1. Missing unit tests for critical data pipeline and mathematical functions (loss, metrics, stitching).
  2. Bug in Keras metrics DiceCoefficient and MeanIoU where add_weight received an invalid shape argument.
  3. Use of A.RandomBrightnessContrast and similar augmentations in dataset.py that destroyed the physical reflectance signatures of clouds and water bodies.
  4. Python GDAL package installation failure due to a mismatch with the system library version.

Technical Implementation

  • Removed A.RandomBrightnessContrast, A.GaussNoise, and A.GaussianBlur from _build_albumentations_pipeline in dataset.py.
  • Fixed DiceCoefficient and MeanIoU in model.py by properly using the name kwargs in add_weight.
  • Downgraded GDAL in requirements.txt to 3.8.4 to align with the system-provided libgdal-dev version.
  • Added a full pytest suite under tests/ covering dataset.py, geospatial_utils.py, and model.py.

Validation (tests, metrics)

  • Added test_dataset.py to ensure CloudPatchDataset correctly reads .npy arrays, maintains batch shapes, and properly instantiates train/val splits.
  • Added test_model.py to mathematically validate MultiClassDiceLoss, CombinedDiceCELoss, DiceCoefficient, and MeanIoU against expected tensor operations.
  • Added test_geospatial.py to validate generate_tile_coords, cosine_bell_mask, and the sliding window stitch_predictions logic.
  • All tests pass (pytest tests/).

Impact

  • Accuracy: Ensures physics-based integrity by removing spectral-altering augmentations.
  • Stability: Prevents runtime crashes from invalid metrics initialization and provides automated verification for future changes.