Convml-tt

[Full Changelog](https://github.com/convml/convml_tt/compare/v0.14.1...v0.14.0)

*maintenance*

- Fix flake8 config so that pre-commit runs [\86](https://github.com/convml/convml-tt/pull/86)

[Full Changelog](https://github.com/convml/convml_tt/compare/v0.14.0...v0.13.1)

*new features*

- New faster grid-based manifold plot which constructs a single image from
individual tile images, rather than plotting individual images as floating
axes.
[\82](https://github.com/convml/convml_tt/pull/82)

- Make it possible to plot grid-overview into a set of axes defined externall
from the function call
[\85](https://github.com/convml/convml_tt/pull/85)

*maintenance*

- Pin to pytorch-lightning < 2.0 and apply fixes for matplotlib 3.8.0 to handle
breaking changes. [\84](https://github.com/convml/convml-tt/pull/88)

[Full Changelog](https://github.com/convml/convml_tt/compare/v0.13.1...v0.13.0)

*bugfixes*

- For image sliding-window (tiling) datasets use the loaded dataframe
describing the tiling to set the tile size
[\80](https://github.com/convml/convml-tt/pull/80)

*maintenance*

- Support pytorch-lightning >1.8, Update README with new cudatoolkit install,
Update setup.cfg to reflect package changes and update pre-commit and linting
tools [\81](https://github.com/convml/convml-tt/pull/81)

[Full Changelog](https://github.com/convml/convml_tt/compare/v0.13.0...v0.12.0)

*new features*

- Add more manifold transform options (LLE, LTSA, Hessian-LLE, MDS, TSNE)
available in `interpretation.embedding_transforms` and add generic manifold
plotting routine `interpretation.plots.manifold2d`. In addition, manifold
plots now use a fraction of the cumulative distribution of anchor-neighbor
tiles to determine which tiles to render
[\76](https://github.com/convml/convml_tt/pull/76)

- All tile-datasets (`convml_tt.data.dataset.ImageTripletDataset`,
`convml_tt.data.dataset.ImageSingleDataset` and
`convml_tt.data.dataset.MovingWindowImageTilingDataset`) now internally use a
`pandas.DataFrame` to hold information about individual source files. This
makes it easier to for example add meta-information into embedding DataArrays
produced from a dataset. [\74](https://github.com/convml/convml_tt/pull/74)
[\77](https://github.com/convml/convml_tt/pull/77)
[\78](https://github.com/convml/convml_tt/pull/78)

*maintenance*

- Update black to v22.3.0
[\75](https://github.com/convml/convml_tt/pull/75)

- Fix example data and pretrained model download issue by avoiding check of SSL
certificate of homepages.see.leeds.ac.uk
[\79](https://github.com/convml/convml-tt/pull/79)

[Full Changelog](https://github.com/convml/convml_tt/compare/v0.11.0...v0.12.0)

*new features*

- Add option for reduced precision training (16bit) which enables training on
hardware with less memory available
[\62](https://github.com/convml/convml_tt/pull/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](https://github.com/convml/convml_tt/pull/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](https://github.com/convml/convml_tt/pull/59)

*breaking changes*

- in line with the rest of the python community (see eg
https://github.com/pydata/xarray/issues/6138) support for python 3.7 has been
dropped [\73](https://github.com/convml/convml_tt/pull/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](https://github.com/convml/convml_tt/pull/67)

*general improvements*

- Option to skip tile triplets with missing tiles when creating `TripletDataset`
[\64](https://github.com/convml/convml_tt/pull/64)

- Improve dendrogram plot to give ability to return tile-clusters as
`xr.DataArray` and annotated scatter plot to support text labels
[\70](https://github.com/convml/convml_tt/pull/70)

*maintenance*

- Ensure loading of fastai v1 models (from weights) also works when we have a
GPU available
[\68](https://github.com/convml/convml_tt/pull/68)

- Add CI action to automatically publish tagged releases on pypi
[\63](https://github.com/convml/convml_tt/pull/63)

- Move all tile and cartesian regridded data generation to a separate package
[convml-data](https://github.com/convml/convml-data)
[\56](https://github.com/convml/convml_tt/pull/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](https://github.com/convml/convml_tt/pull/60)

- Move examples downloading functionality into package instead of relying on
torchvision since the code in there keeps changing and breaking our examples
[\61](https://github.com/convml/convml_tt/pull/61)

[Full Changelog](https://github.com/convml/convml_tt/compare/v0.10.1...v0.11.0)

*new features*

- Annotated scatterplots now work for `data.dataset.MovingWindowImageTilingDataset`s
[\52](https://github.com/convml/convml_tt/pull/52)

- `utils.get_embeddings` now uses GPU for producing the embeddings if one is
available, drastically speeding up inference time.
[\53](https://github.com/convml/convml_tt/pull/53)

*breaking changes*

- `utils.make_sliding_tile_model_predictions` has been removed in favour of
using `utils.get_embeddings` directly
[\54](https://github.com/convml/convml_tt/pull/54)