v0.0.6: support for non-binary treatment

Update pypsps to support non-binary treatment and generic outcome handling.

in particular, allows users to specify any y_pred and y_true for outcome and treatment, as long as it fits into k columns in the overall y_true and y_pred of the inputs/outputs of a pypsps model.

v0.0.5

• fix relative imports
• add default loss functions

v0.0.4 - configurable model architecture, support any distribution loss

• more than just toy model architecture supported for binary treatment and normal distribution outcome
• allows for any distribution based loss function as outcome loss using tfp/distributions
• misc cleanup and unit test updates

v0.0.3 - poetry package mgmt; workflows

• changed to poetry package mgmt
• added python Github workflows for PRs

v0.0.2 - cleanup, additions, fixes

What's Changed

• Cleanup of loss fcts; adding new metrics; serialization; code cleanup by @gmgeorg in #1

  change order of split_y_pred() to always have propensity score as last entry ([-1])
  add a propensity score AUC metric (inherting from tf.keras.metrics.AUC() class)
  add aggregation function for outcome predictions based on weights of each state
  fix unit tests

Initial release

Very first release of pypsps.

Has relevant implementation to replicate work in the PSPS paper.