Fixed * removed use of deprecated sklearn externals
0.5.8
Fixed: * `CategoryTree.decision` now always refines categories with a single parent * `run_tests.py` now correctly returns exit code.
Added * `batch_convert_to_cog` to `util_gdal` * `batch_validate_cog` to `util_gdal`
0.5.6
Added * `util_lru` containing implementations of a dictionary based LRU cache.
Changed * AbstractFrames now uses the configurable `LRUDict` instead of the `functools` decorator. * Moved toydata from ndsampler to kwcoco
0.5.5
Changed * Moved `ndsampler.util_futures` and `ndsampler.util` into `ndsampler.utils`. Old ways of accessing these files will raise deprecation warnings.
Fixed: * Using a better stratified split algorithm * Fixed issue with prepare and the npy backend when too many files were opened
Added
* Added `boxsize_stats` to `CocoDataset`.
0.5.4
Fixed
* `SerialFuture` was broken in python 3.8 * Fixed incorrect use of the word "logit", what I was calling logits are actually log probabilities.
Added
* `_ensure_imgsize` can now use threads
Changed
* `CocoDataset.union` now preserves image ids when possible.