Skip to content

Latest commit

 

History

History
 
 

releases

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

Versions

V0.4 version 31/10/2017

  1. Added original libfm, libffm classification and vowpal wabbit wrappers. 5 algorithms in total.
  2. Added input_index command. This allows StackNet to be run with user-provided indices as a separate file.
  3. Added include_target command. It appends the target variable in the beginning of the file when output_name is used.
  4. Added the ability to make comments in the params files using #. Anything on the right of this symbol is regarded as comment.
  5. Fixed an assertion error in SklearnknnClassifier.
  6. Fixed bug in scoring for (StackNet's implementation of) libfm.
  7. Fixed other minor errors.

V0.3 version 06/08/2017

  1. Added Python subprocesses. The user needs to install python himself/herself and have python available on PATH
  2. Added 22 new algorthms in total (Regerssors and Classifiers)
  3. Added various algorithms based on Sklearn compatible with version 0.18.2. Specifically : SklearnAdaBoostClassifier, SklearnAdaBoostRegressor, SklearnDecisionTreeClassifier, SklearnDecisionTreeRegressor, SklearnExtraTreesClassifier, SklearnExtraTreesRegressor, SklearnknnClassifier, SklearnknnRegressor, SklearnMLPClassifier, SklearnMLPRegressor, SklearnRandomForestClassifier, SklearnRandomForestRegressor, SklearnSGDClassifier, SklearnSGDRegressor, SklearnsvmClassifier, SklearnsvmRegressor
  4. Added support for keras' algorithms through python, compatible with version 2.0.6. It was tested with tehano 0.9.0, but it should work with tf too. The user is responsible for installing keras and for optimizing its backend (and make sure it is available through python). Specifically added KerasnnRegressor or KerasnnClassifier.
  5. Added support for user-defined python scripts. The user can name them as PythonGenericRegressor[index] or PythonGenericClassifier[index] and put them inside lib/python (see PythonGenericClassifier0 example to understand the right structure). He/She could then call it within the parameters as PythonGenericClassifier index:0.
  6. Added Fast_rgf and sepcifically FRGFClassifier and FRGFRegressor
  7. Fixed a back for failling to cast StackNetClassifier as StackNetRegressor when task=predict.
  8. Added the display of average metric for all models at the end of each level (eg. average logloss of all folds for each model)

V0.2 version 25/06/2017

  1. Added bagging as ab extra hyper parameter in each model . It could be specified as bags:3
  2. Added (compiled) lightGBM and created an LightgbmClassifier and LightgbmRegressor based on StackNet's api
  3. Added H2O-3's Algorithms
  4. Specifically added H2OGbmRegressor, H2ODeepLearningRegressor ,H2ODrfRegressor, H2OGlmRegressor
  5. and H2OGbmClassifier, H2ODeepLearningClassifier ,H2ODrfClassifier, H2OGlmClassifier, H2ONaiveBayesClassifier.
  6. Fixed a bug in bins that was causing one less bin to be created
  7. Added BaggingClassifier and BaggingRegressor

V0.1 version 09/06/2017

  1. Added regression which can be defined with a new task parameter in the command line. Otherwise the user needs to set classification. Task is now a mndatory parameter
  2. Added (compiled) Xgboost and created an XgboostClassifier and XgboostRegressor based on StackNet's api
  3. Fixed baug with lowercase true being ignored from parameters
  4. Fixed bug that was not ignoring zeros loaded from sparse data
  5. Added an equalsizebinner in preprocess.binning
  6. Added a bins command in the train argument that that allows classifiers to be used in regression problems. The target variable gets binned using equalsizebinner and then used in classification.
  7. Added an rsquared in metrics

V0.0 version 01/04/2017

  1. Made significant changes to the ml.tree module. Previously it was built assuming a dense format. Whereas now it is built having a sparse format at its base. Significant changes have been made to enable hashing to be more efficient (but with a little bit more memory). Dense format might have suffered a bit speed-wise, but it will be improved in the future.
  2. Improved multithreading in ml.tree
  3. Included a indices_name in the command line to print a .csv file for each fold with the corresponding train(0) and valiation(1) indices one stack under the other .The format it row_index,[0 if training else 1]. First it prints the train indices and then the validation indices in exactly the same order as they appear when modelling inside StackNet.
  4. Fixed a bug that was printing the same prediction file twice int he last level of the training process.
  5. Fixed a bug when there was not a target variable (or an extra column) in the test file
  6. Fixed a bug with reading data in Sparse format
  7. Added standardscaler in code, BUT it is not available in StackNet yet (only in source code)