Hackathon Framework

Objective of this project to enable quick experimentation in Data Analytics projects with minimal cookie cutter programming. Getting rid of all the fit_transforms.!

---

NOTE

• This is a work in progress. Underlying modules are in process of development.
• As this project matures there will be changes in the scripts such as train.py and predict.py
• TODO
  • Create modules for tuning, stacking
  • Removal of some of the modules that are redundant

---

The framework is designed to make the Data Science flow easier to perform, by encapsulating different techniques for each step within 1 method. There are classes for each of the below listed steps:

• Feature Evaluation

  • Report to give an intution of the dataset

• Feature Engineering

  • Modules to perform feature transformations on Categorical and Numerical Dataset.
  • Various applicable techniques are encoded within these modules and are accesed with an argument.

• Fefature Generation

  • Module to create new features based on different techniques

• Cross Validation

  • Stratified Folding both for Regression and Classification

• Training

  • Run multiple models using 1 class.
  • Evaluating and Saving the results in an organized manner

• Tuning

  • Hyper-parameter tuning of multiple models, based on json arguments for parameter values.

• Prediction

• Evaluating the model

Steps to use the framework

1. Clone the repo.
2. Create 3 folders input and model and tuneq.
3. Save the training, testing and sample submission file in input folder.
4. The outputs generated from training such as trained model, encoders and oof_preds will be saved in model folder.
5. The parameters for fine tuning the models should be saved in the tune folder.
6. Update the config.py to point it to the correct path for data, model and tuning.
7. Update the dispatcher.py with model/models you want to run your dataset on.
8. Use the sample notebook to understand how to use this framework after this intial configuration is completed.

Description of Files and their Purpose

• config.py: Config file to give path of all the datasets and other standard configuration items. Such as csv files path, random seed etc.

• feature_eval.py: This script and the class inside is used to analyze the dataframe and its columns to get the following output:

  • min, max and unique values of each column
  • histogram/ distribution of each column
  • corelation of columns using a heat map

• feature_gen.py: Encapsulates method to generate new features. Currently implemented the Polynomial features method from sklearn. Returns Dataframe with new features.

• feature_impute.py: Encapsulates the method to impute blank values in a dataframe. Currently, it supports 3 imputation methods:

  • Simple Imputer
  • Model Based Imputer: Extra Trees or knn
  • Knn based imputer
  • Returns updated Dataframe

• cross_validation.py: This class is used to perform cross validation on any dataframe based on the type of problem statement. It is used to create cross validated dataset.

• categorical.py: This class can be used for encoding of categorical features in a given dataframe.

  • Inputs : Dataframe, Categorical Columns List, Type of Encoding
  • Output: Encoded Dataframe
  • Supported Encoding Techniques:
    • Lable Encoding
    • Binary Encoding
    • One Hot Encoding

• numerical.py: This class can be used for encoding of numerical features in a given dataframe.

  • Inputs : Dataframe, Categorical Columns List, Type of Encoding
  • Output: Encoded Dataframe, Transformer Object for later use.
  • Support Techniques:
    • Standard Scaler
    • Min-Max Scaler
    • Power Tranformer
    • Log Transformer

• metrics.py: This class can be used to evaluate the results of given predictions and actual value.

• dispatcher.py: Python File with Models and parameters. They have been designed to supply the models to engine.py for training on a given dataset

• engine.py: This script encapsulates the method to train and evaluate the multiple models simultaneously

  • Leverages on dispatcher.py and metrics.py for model and metrics
  • The results for each fold are also saved in the models folder as oof_predictions.csv for each model.
  • To Do Stacking module to suporrt stacking of multiple models

• Scripts to be ignored for now:

  • train.py: For training
  • predict.py: For prediction
  • tune.py: For tuning h-parameter
  • create_folds.py: To create folded datframe