v0.12.0

Full Changelog

new features

• Add option for reduced precision training (16bit) which enables training on
  hardware with less memory available
  #62

• Add new isomap 2D embedding manifold plot which takes the best (nearest
  anchor-neighbor) triplets and plots these at a uniform sampling across an
  isomap 2D extracted embedding manifold
  #57

• annotated scatterplot now uses matplotlib's AnnotationBbox which means that
  annotated scatterplots can be rescaled (with say ax.set_xlim) with
  annotations staying in place and keep a fixed size
  #58

breaking changes

• in line with the rest of the python community (see eg
  pydata/xarray#6138) support for python 3.7 has been
  dropped #73

bugfixes

• Fix bug in dendrogram plot where wrong tiles were shown when creating a
  dendrogram from a data-array of embedding vectors that have been filtered (so
  that the tile ids don't just from 0, 1, 2 etc)
  #67

general improvements

• Option to skip tile triplets with missing tiles when creating TripletDataset
  #64

• Improve dendrogram plot to give ability to return tile-clusters as
  xr.DataArray and annotated scatter plot to support text labels
  #70

maintenance

• Ensure loading of fastai v1 models (from weights) also works when we have a
  GPU available
  #68

• Add CI action to automatically publish tagged releases on pypi
  #63

• Move all tile and cartesian regridded data generation to a separate package
  convml-data
  #56

• Fix CI testing by switching to microconda and unpinning pytorch version
  (since kornia now supports pytorch==1.8.0) and switch to pre-commit for
  linting while fixing linting issues picked up during this switch.
  #60

• Move examples downloading functionality into package instead of relying on
  torchvision since the code in there keeps changing and breaking our examples
  #61