v0.3.6

What's Changed

• Support non-Gaussian ConvNP likelihoods in low-level and high-level prediction interfaces by @tom-andersson in #97
• Fix multiple N-D target sets case in ConvNP; Update ConvNP API
• Minor suggestions to installation docs. by @davidwilby in #98

New Contributors

• @davidwilby made their first contribution in #98

Full Changelog: v0.3.5...v0.3.6

v0.3.5

What's changed

• Fix bug in batchwise Trainer training with multi-dimensional context sets (#93; thanks @nilsleh)

Full Changelog: v0.3.4...v0.3.5

v0.3.4

• Fix PyPI upload by removing get-station-data dependency; instead raise error that it must be installed manually if user calls the function that uses it

Full Changelog: v0.3.3...v0.3.4

v0.3.3

What's Changed

• New documentation by @tom-andersson and @kallewesterling, containing steps for getting started, a user guide, learning resources, a list of research ideas, community information, and more. Uses the Jupyter Book system with fully reproducible notebooks demoing features of the package.
• New deepsensor.data.sources module for downloading ERA5, GHCNd, topography, and land mask data directly into DeepSensor xarray/pandas format. Results can be cached to disk to avoid re-downloading. Leverages get-station-data tool by @scotthosking, @magnusross, and @tom-andersson.
• New plotting tools: deepsensor.plot.prediction and deepsensor.plot.task.
• Various bug fixes.

Breaking changes

• The ConvNP hyperparameter points_per_unit has been renamed to internal_density to be more intuitive. Instances of points_per_unit in code and model JSON configuration files will need to be renamed to internal_density.

New Contributors

• @scotthosking made their first contribution in #86

Full Changelog: v0.3.2...v0.3.3

v0.3.2

• Fix bug in ContextDist acquisition function (thanks @acocac)

v0.3.1

What's Changed

• Fix bug in ContextDist acquisition function (thanks @acocac)

Full Changelog: v0.3.0...v0.3.1

v0.3.0

What's Changed

• New Prediction object output by model.predict, containing xarray or pandas data (#53)
• New TaskLoader functionality for generating satellite gap-filling training tasks (#23, example)
• DataProcessor scales each spatial dimension equally to fix stripe artefacts, and raises a warning if user-provided coord mappings don't do this (#77)
• Task operations now work like task.operation_method() and are tracked in the task["ops"] list
• TaskLoader does not sample target data for the Task if target_sampling isn't provided (supports context-only Tasks for inference with model.predict)
• Support spatiotemporal aux_at_targets xarray data (with a time dim), not spatial-only
• Fix model.predict not permitting numerical noise in X_t coordinates (#78)
• Key classes are now exposed from higher import levels (de876d4)
• Add CONTRIBUTING.md and CODE_OF_CONDUCT.md by @kallewesterling
• Issue templates by @kallewesterling
• Google docstrings added to the majority of the package by @kallewesterling
• DeepSensor Slack channel request form
• Preliminary sphinx readthedocs page by @kallewesterling

Breaking changes

• Replace any mean_ds, pred_ds, samples_ds = model.predict(...) with pred = model.predict(...). Prediction pandas/xarray objects can now be accessed with pred["<var_ID>"].
• Replace any remove_nans_from_task_Y_t_if_present(task) with task.remove_target_nans()
• DataProcessor now auto-normalises coordinates differently to preserve aspect ratio - see #77. If you aren't specifying spatial coordinate normalisation mapping explicitly (through x1_map/x2_map or through config), your trained models may start receiving different data.

Full Changelog: v0.2.5...v0.3.0

v0.2.5

What's Changed

• Fix incompatibility with Python < 3.10 (thanks @patel-zeel)

Full Changelog: v0.2.4...v0.2.5

v0.2.4

What's Changed

• Modularise and track Task operations (eg adding batch dim, removing nans, converting to tensor).
• Fix bug in concat_tasks when there are NaNs in target data (thanks @nilsleh!)

Full Changelog: v0.2.3...v0.2.4

v0.2.3

What's Changed

• Bug fixes

Full Changelog: v0.2.2...v0.2.3