v0.1.9

What's Changed

• Fix bug in TaskLoader.load_dask (thanks @magnusross)
• Handle dates with no pairs of stations in compute_pandas_data_resolution

Breaking changes

• The TaskLoader.__init__ aux_at_contexts arg is no longer a tuple of (int, xarray.DataArray), and is now just an xarray object that will be automatically sampled at the off-grid context locations from all context sets.

Full Changelog: v0.1.8...v0.1.9

v0.1.8

What's Changed

• Compute analogue of resolution for scattered pandas data for inferring default ConvNP internal grid resolution
• Allow sampling auxiliary data at locations of another context set in TaskLoader with aux_at_contexts kwarg in TaskLoader.__init__
• deepsensor.plot.offgrid_context_observations utility to plot context observation numerical values
• Fix bug in concatenating/batching Task objects with gridded data (thanks @patel-zeel)
• Support active learning with aux_at_targets downscaling functionality by @RohitRathore1 in #35
• Unified (single) progress bar during active learning with GreedyAlgorithm

Breaking changes

• Update DataProcessor config to store user-provided normalisation method for each variable

Contributors

Full Changelog: v0.1.7...v0.1.8

v0.1.7

What's Changed

• Minor bugfixes in deepsensor.active_learning.acquisition_fn

Full Changelog: v0.1.6...v0.1.7

v0.1.6

What's Changed

• Downscaling using auxiliary output MLP
• Assign maximisation/minimisation attr to each AcquisitionFunction so the user doesn't need to decide (thanks @jonas-scholz123 @polpel)

Full Changelog: v0.1.5...v0.1.6

v0.1.5

What's Changed

• Fix rounding errors in DeepSensorModel.predict coordinates from normalise-unnormalise operations by @tom-andersson and @polpel in #25
• Support autoregressive (AR) sampling in DeepSensorModel.predict
• Support training with multiple non-overapping targets
• Support NaNs in context and target data for ConvNP
• deepsensor.active_learning enhancements
• Support numpy coordinates in TaskLoader context/target and DeepSensorModel.predict targets
• Switch default random sampling behaviour for xarray data to use linear interpolation rather than grid-cell wise to avoid prediction artifacts away from grid cells (thanks @jonas-scholz123)

Full Changelog: v0.1.4...v0.1.5

v0.1.4

• Fix decoder scale not being inferred from model discretisation density (ppu). Closes #18
• Add deepsensor.plot.feature_maps method
• Provisional deepsensor.active_learning functionality with acquisition functions and greedy algorithm

Breaking changes

• ConvNP class moved from deepsensor.model.models to deepsensor.model.convnp

Full Changelog: v0.1.3...v0.1.4

v0.1.3

Try to trigger PyPI upload with publisher set up on PyPI

Full Changelog: v0.1.2...v0.1.3

v0.1.2

What's New

• Breaking change: Plotting module moved and methods renamed for nicer-looking imports. E.g. now deepsensor.plot.plot.plot_context_encoding is deepsensor.plot.context_encoding.
• Added deepsensor.plot.receptive_field method for plotting model's RF in unnormalised space with cartopy coastlines.

Full Changelog: v0.1.1...v0.1.2

v0.1.1

What's New

• Published on PyPI
• ConvNP class defaults to CNP model
• More intuitive names for ConvCNP or ConvCNP: ConvNP(..., likelihood="cnp") and ConvNP(..., likelihood="gnp")

Full Changelog: v0.1.0...v0.1.1

v0.1.0

Pre-release with basic DeepSensor functionality for data processing, task loading, and neural process training/inference.

What's New

• DataProcessor for normalising xarray and pandas data + standardising coordinates
• TaskLoader for loading neural process meta-learning tasks from xarray and/or pandas data, outputting Task objects
• TaskLoader.__call__ provides sampling schemes for generating context and target sets. Options:
  • random sampling (xarray/pandas),
  • passing all observations (xarray/pandas),
  • randomly splitting into context & target (pandas only).
• ProbabilisticModel class providing blueprint for generic model interface
• DeepSensorModel(ProbabilisticModel) class for outputting unnormalised model predictions in xarray (grid) or pandas (off-grid)
• ConvNP(DeepSensorModel) model class wrapping around neuralprocesses (https://github.com/wesselb/neuralprocesses) for convolutional neural process modelling
• train_epoch method implementing simple training scheme on a list of Tasks

Contributors

• Thanks to @wesselb for support with backend-agnosticism!
• DataProcessor dimension validation + unit tests by @jonas-scholz123 in #2
• fix str of tensorflow backend by @acocac in #3
• Fix else level in set_gpu_default_device() by @polpel in #4
• Fix DataProcessor's validaiton of dimension ordering in xr.Dataset by @polpel in #5

Full Changelog: https://github.com/tom-andersson/deepsensor/commits/v0.1.0