Galaxy-ml

New Features

- Adds circleci config for both api and tool tests.
- Adds train_test_split tool which supports shufflesplit, stratifiedshufflesplit, groupshufflesplit and orderedtarget split.
- Adds fitted_model_eval tool.
- Refactors binarize target estimators. There are a lot of improvements. One of them is that the estimator family now support most sklearn scorers.
- Adds clean_params in utils
- Adds cv_results_ outputs for nested inner CV and unfitted searchCV object from searchCV tool.
- Adds keras training and evaluation tool.
- Adds support of decision_function for binarize target classifiers.
- Adds matplotlib svg format option in `ml_visualization_ex` tool.
- Adds 'sklearn.ensemble.HistGradientBoostingClassifier' and 'sklearn.ensemble.HistGradientBoostingRegressor'
- Adds new regression scorer `max_error`.
- Upgade scikit-lean to v0.21.3, mlxtend to v0.17.0, imbalanced-learn to v0.5.0, keras to v2.3.1 and tensorflow to v1.15.0.

Changes

- Replaces all generators' `fit` with `set_processing_attrs`.
- Raise ValueError instead of [0, 1] normalization when predictions from `BinarizeTargetRegressor` go out of range.
- Refactor `iraps_classifier` module. Binarize target estimators do the same prediction as the wrapped estimator. A delicated `predict_score` is made to work with binarize scorers.
- Changes precision-recall curve and ROC curve to take headers and upgrade plotly to v4.3.0 in ml_visualization_ex tool
- Change to dynamic output of pipeline or final main estimator

Bug Fixes

- Fixes random_state error in `_predict_generator`.
- Fixes stale path issue by replace relative paths with full paths.

New Features

- Adds searchcv tools to output `weights` for deep learning models.
- Makes `KerasGBatchClassifier.evalue` to support multi-class and multi-label classification problem.
- Adds parameter `verbose` in KerasG models to output device placement.
- Adds `metrics` in keras model building tools.
- Makes `train_test_eval` tool.
- Makes `GenomicVariantBatchGenerator`.
- Makes `model_prediction` tool to support `vcf` file type.
- Adds plotly plotting tool facility for `feature_importances`, `learning_curve`, `pr_curve` and `roc_curve`.
- Adds `_predict_generator` to output y_true together with prediction results.
- Adds support of `return_train_score` for `KerasGBatchClassifier` in gridsearchcv.
- Adds `ml_visualization.xml` tool support many plots.

Changes

- Changes dependency `tensorflow` to `tensorflow-gpu`.
- Moves all tools to folder `tools`.
- Makes `sklearn.preprocessing.Imputer` deprecated.
- Updates dependencies in `requrements.txt`.
- Refactor `keras_model_config` tool by grouping layer key words arguments.
- Refactor the `preprocessors.py` into folder structure.

Bug Fixes

- Fixes `KerasGBatchClassifier` doesn't work with callbacks.
- Fixes `GenomicIntervalBatchGenerator` doesn't work in nested model validation.
- Fixes `GenomicIntervalBatchGenerator` failed for sequences in blacklist matches.

New Features

- Adds MIT license.
- Adds `setup.py` and `requirement.txt` for APIs installation.
- Makes Galaxy-ML APIs as a library and installable vis pypi and bioconda.
- Adds `GenomicIntervalBatchGenerator`, an online data generator that provides online genomic sequences transformation from a reference genome and intervals. By trying to offer the same functionalities of [selene](https://github.com/FunctionLab/selene), `GenomicIntervalBatchGenerator` is implemented by, 1) reusing selene cython backend; 2) extending `keras.utils.Sequence`, multiple processing and queueing capable; 3) compatibilizing with sciKit-learn APIs, like KFold, GridSeearchCV, _etc_. `GenomicIntervalBatchGenerator` is supposed to be fast and memory-efficient.
- Adds parameter `steps_per_epoch`, `validation_steps` to `BaseKerasModel`.
- Adds parameter `prediction_steps` to `KerasGBatchClassifier`.
- Adds `class_weight`-like parameter `class_positive_factor` to `KerasGBatchGenerator` for imbalanced training.


Changes

- Refactor fast array generators, introduced `fit` method.
- Refactor iraps_classifier random index generator, reduce fit time by about 45%

Bug Fixes

New Features

- Adds `validation_data` into keras galaxy models and supports gridsearch and `model_validation`.
- Adds fasta sequence batch generator and makes `FastaDNABatchGenerator`, `FastaRNABatchGenerator` and `FastaProteinBatchGenerator`.
- Adds keras galaxy batch classifier and `generator.flow`.
- Adds `keras_batch_models` tool.
- Adds `GenomeOneHotEncoder` and `ProteinOneHotEncoder` to `pipeline`.
- Adds API documentation at `https://goeckslab.github.io/Galaxy-ML/`.
- Extends `BinarizeTarget Classifier/Regressor` to support `fit_params`.
- Modifies `read_columns` function to avoid repeated input file reading.

Changes

- Changes `model_validation` tool name.


Bug Fixes

- Fixes CV groups file issue in searchcv tool.
- Fixes cheetah error in model_validation tool.

New Features

- Adds `BinarizeTargetTransformer`.
- Adds support of binarize_scorers to `BaseSearchCV`.
- Adds `sklearn.ensemble.VotingClassifier` and `VotingRegressor` (will be available sklearn v0.21).
- Enhances security of `try_get_attr` by adding `check_def` argument.
- Adds `__all__` attribute together with `try_get_attr` to manage custom module and names.
- Adds keras callbacks. Now supports `EarlyStopping`, `RemoteMonitor`, `TerminateOnNaN`, `ReduceLROnPlateau` and partially support `ModelCheckpoint`, `CSVLogger`.

Changes

- Pumps `stacking_ensembles` too to version 0.2.0.
- Changes `KerasBatchClassifier` to `KerasGBatchClassifier`.

Bug Fixes

- Fix voting estimators duplicate naming problem.

New Features

- Adds Nested CV to searchcv tool.
- Adds `BinarizeTargetClassifier`.classifier_, `BinarizeTargetRegressor`.regressor_ and `IRAPSClassifier`.get_signature() in estimator_attributes tool.
- Reformat the output of `corss_validate`.
- Adds `KerasBatchClassifier`.
- Makes `KerasGClassifier` and `KerasGRegressor` support multi-dimension array.

Changes

- Changes min value of `n_splits` from 2 to 1.
- Main Tool version changes on the last second number instead of the last one.

Bug Fixes

- Fixes `train_test_split` which didn't work with `default` scoring.