Nyoka

Bug Fix
* Bug fix for ArimaToPMML class (`AttributeError` when DataFrame is used for training)

Major Update
1. Added exporter for Object Dectection Model (RetinaNet)
2. Added extra parameter to Keras exporter to add python script export functionality.
3. Reduced list of parameters in ARIMA exporter

Bug Fix
1. Bug fix for Seasonal ARIMA (Residual for Seasonal Component)
2. Bug fix for AnomalyDetectionModel (Calculation of the threshold)
3. Bug fix for NeuralNetwork models

Minor update
* PMML representaion updated
* AnomalyDetectionModel is used for IsolationForest instead of MiningModel
* stddev as a build-in function
* Removed threshold from OutputField

Bug fix
* Changed operator in SimplePredicate for LightGBM
* Changed default wordSeparatorCharacterRE in TextIndex from `(?u)\b\w\w+\b` to `\s+`
* Corrected invalid constantTerm, TimeValue in ARIMA.
* Added RMSE, Residuals to ARIMA

Bug fix
* Output dataType is updated to have same data type as the classes have.
* batchNormalizationCenter is added to LayerParameters
* Removed unused PMML elements generated in case of LabelBinarizer and OneHotEncoder
* feature_importances_ not found in IsolationForest

Minor Update
* Added pre-processing Transformer `Lag`
Bug Fix
* OneHotEncoder has not attribute `_active_features_`
* Parameters for `ArimaToPmml` and `ExponentialSmoothingToPmml` have made optional. Now users need to pass only `result_objs` to the exporter. No need to pass `time_series_data` and `model_obj`

Change in PMML representation
* Logistic Regression
* Ridge Classifier
* SGD Classifier
* Linear SVC
* Linear Discriminant Analysis
Bug fix
* Linear SVC does not support probability
* KMeans Comparison Measure
* PCA, Binarizer, Gradient Boosting Classifier, Gradient Boosting Regressor change in precision
* CountVectorizer generates unsupported bytes.
* Removed duplicate OutputField