Madminer

New features:
- Morphing-aware likelihood ratio estimators. See https://arxiv.org/abs/1805.00020 for a description. Implemented in `madminer.ml.MorphingAwareRatioEstimator`.
- Gradient clipping can now be set with the keyword `clip_gradient` in the estimator `train()` functions.

New features:
- Reweight existing samples (either generated with MadMiner or standalone MadGraph) through `MadMiner.reweight_existing_sample()`
- Custom parameter grids / evaluation points in `AsymptoticLimits` through the new keyword `thetas_eval`

New features:
- Dropout support
- Many more activation functions
- Number of workers for data loading can be specified

Bug fixes:
- Fixed crash in DelphesReader and LHEReader when no systematics are used
- Fixed logging error with double parameterized ratio estimation methods

New features:
- Expanded systematics system. Users now declare systematics with `MadMiner.add_systematics()`, which in addition to the previous PDF and scale variations also allows normalization uncertainties. When adding samples in the `LHEReader` and `DelphesReader` functions, each sample can be linked to an arbitrary subset of systematics, giving the user a lot of flexibility.

Breaking / API changes:
- For processes with systematic uncertainties, the MadMiner file format changed in a not-backward-compatible way. Please do not use files with systematic uncertainties that were generated with MadMiner versions before v0.6.0 with the new code version (MadMiner will crash). Sorry about this.

Bug fixes:
- Fixed wrongly including the central element in the calculation of the PDF uncertainties.

Documentation:
- Updated and expanded tutorial on systematic uncertainties.

Internal changes:
- Some internal changes related to nuisance parameters, including the MadMiner file format.

New features:
- Automatic shuffling of MadMiner HDF5 files after reading in LHE or Delphes files

Bug fixes:
- Fixed rare crash in `AsymptoticLimits`

New features:
- Clean separation between training and validation events: the `SampleAugmenter` functions have new keywords `partition` and `validation_split`. With `partition="validation"`, validation data without potential overlap with the training samples can be generated. In the `madminer.ml` classes, this can be provided with new keywords like `x_val`, `theta_val` when calling `train()`.
- More consistent calculation of the Fisher information covariance: the covariance matrices in `mode="score"` are now the ensemble covariance, without dividing by `sqrt(n_estimators)` as before. This is also the default behavior. The old default behavior can be used with `mode="modified_score"`.
- Parameter rescaling also for `DoubleParameterizedRatioEstimator` and `LikelihoodEstimator`.
- When continuing the training of a pre-trained model, the parameter and observable rescaling is not overwritten.
- SALLY limits can now be calculated with `Ensemble` instances of multiple score estimators.

Breaking / API changes:
- The `SampleAugmenter` functions do no longer accept the keyword `switch_train_test_events`. Use `partition="train"` or `partition="test"` instead.

Bug fixes:
- Fixed bug in the logging output about relative cross-section uncertainties during the Fisher information calculation.
- Fixed MadMiner crashing when calculating adaptive histogram binnings.
- Fixed bug in `AsymptoticLimits` where only 1 observed event was returned.

Internal changes:
- Abstracted `Estimator` classes with parameter rescaling into new `ConditionalEstimator` class.