GitLab

Add optional cache usage of object properties in `SegmentationDataset` via `cache_properties` kwarg (see docs/api.md)

* Add first example to api.md: Usage of the SegmentationDataset property caches
* Property cache can be passed on to `SuperSegmentationDataset` via `sd_lookup`, which is also available for `SuperSegmentationObject` retrieved through `SuperSegmentationDataset.get_supersegmentation_object` (see docs/api.md)
* code coverage is now computed correctly for the master branch tests by using a specific pytest config (full_run.ini)
* continued: code adaption to line width of 120

ignore missing config.yml warning in test_models_e3.py; install open3d via conda; exclude tests from setup.py packages

Multi-view pipeline refactoring and object segmentation feature

See merge request pschuber/SyConn!34

fix scaling in batchjob for sample location caching in case of SegmentationObjects with zero vertices

* add cython method for relabeling 3D arrays based on an ID->ID map
* make per-chunk object IDs in `_gauss_threshold_connected_components` contiguous when filtering small object seeds during watershed
* add `min_seed_vx` parameter for minimum watershed seed size
* code clean-up

improve cache re-usage and performance of loading SuperSegmentationObject meshes ('sv', 'sj', 'syn_ssv', ..); ssv.load_mesh('syn_ssv') and subsequent typedsyns2mesh call in `map_synssv_objects_thread` re-use meshes of 'syn_ssv' objects

refactor meshing - now sparse and dense meshing are based on zmesh and use the same parameters; add model and loader kwarg parser method

- add on the fly view rendering for spines, cell types and morphology embedding
- remove support for cell type inference with large views
- improve speed and quality of meshing (isotropic downsampling)
- add a more flexible image processing during object extraction (e.g. morphology operation and watershed)
- remove unused code
- implement initial frame work for point based models

TODO:
- generate syn_ssv meshes densely, based on a yet to be implemented syn_ssv segmentation KD
- use syn_ssv channel instead of sj for views
- on the fly rendering for glia separation
- support point based models

Conflicts:
	tests/test_run.py

* add bulk loader for meshes of SegmentationObjects
* [WIP] add glia loader for inference
* improve speed of synapse object mapping
* refactored view rendering of large neurons

* improve dataset_analysis speed by better data handling.
* write syconn version during installation and use it in data set config files.
* fix #29 freeze issue which appeared during rendering in environments (environment.yml) recently built from scratch; caused by sklearn.decompostion.PCA called multiple times within short time intervals; fixed by using a custom numpy PCA implementation (`_calc_pca_components` in meshes.py).
* change dependency restrictions.
*
* [WIP] remove elektronn2 model support.
* [WIP] transition to pytorch script, currently missing: sub-cellular structures (in training), mylein (in training).
* [WIP] collect results of dataset_analysis with batch jobs via SLURM.

* add variation to number of target points in glia prediction training
* improve `combine_and_split_syn` by ~8fold with the use of 2-stage kd-trees