The problem.py file

The problem.py file uses building blocks from the RAMP workflow library. These building blocks allow the problem.py file to be relatively simple because the complexity is hidden by the implementation of the building blocks in RAMP workflow. Titanic survival classification challenge will be used as an example when discussing each aspect of the problem.py file. It is worth taking a look at the whole file for reference.

  1. The problem.py file begins by importing any libraries required. For example, the required libraries for the Titanic challenge were:

    import os
    import pandas as pd
    import rampwf as rw
    from sklearn.model_selection import StratifiedShuffleSplit
    
  2. Provide a title:

    problem_title = 'Titanic survival classification'
    
  3. Select a prediction type Prediction types are classes used internally to store predictions and calculate scores. There are different prediction classes to be used with different types of challenges. This is implemented using functions, each of which returns a unique prediction class. The functions currently available are detailed below:

    • make_regression() - used for regression type challenges. Predictions can be 1-dimensional or 2-dimensional for multi-target regression. For multi-target regression a list of ‘target names’ needs to be provided to the parameter label_names.

    • make_multiclass() - used for classification type challenges. Predictions are expected to be 2-dimensional (n samples * n classes) and may be labels or probabilities of each class. A list of label names needs to be provided to the parameter label_names.

  • make_combined() - used for challenges where greater than one type of prediction is made. For example, the drug spectra challenge comprises of a classification task to predict the type of molecule and a regression task to predict the concentration of the molecule.

  • make_clustering() - used for clustering challenges. Prediction should be 2-dimensional (n samples * 2), where one column identifies the samples and the second column identifies the cluster it belongs to. This prediction class was used for the High-energy physics tracking challenge.

  • make_detection() - a unique class specifically designed for the Mars crater detection challenge. A unique algorithm is implemented for combining predictions from different models. See the source code for more information.

Select the appropriate prediction class for your challenge and state this in the problem.py file. If the appropriate prediction class does not exist in RAMP workflow, you can define your own prediction class within the problem.py file. If it is not too specific, we would also encourage you to add your class to RAMP workflow so others can use it in future. See Contributing.

For example, the Titanic survival classification challenge aimed to predict whether or not each passenger survived. Survival is indicated by 0 (did not survive) or 1 (survived). Since this is a classification task the prediction function make_multiclass() is used. Note that by convention label names are stored as a variable _prediction_label_names. The relevant parts of the problem.py file are shown below:

_prediction_label_names = [0, 1]

Predictions = rw.prediction_types.make_multiclass(
label_names=_prediction_label_names)
  1. Select a workflow Workflows implement the steps of the machine learning problem. Each workflow is a class with a train_submission() and a test_submission() function, which defines the workflow to implement during training and testing time. An attribute called workflow_element_names is also required. This attribute is a list of the file names that ramp_test_submission expects for each submission. This class is implemented by RAMP workflow internals to train and test each submission of a challenge. The workflows available are described below:

    • Regressor() - this workflow is for simple regression problems. It will import the file named regressor.py from the submission directory (e.g. submissions/starting_kit/). This regressor.py file should define a class named Regressor that has a fit() and a predict() method. This workflow will run fit() on the training data then at testing time, run predict() on the testing data, using the trained model. If you are using a scikit-learn function, you can simply call the fit() and predict() methods of the model you are using.

  • Classifier() - this workflow is for simple classification problems. It will import the file named classifier.py from the submission directory (e.g. submissions/starting_kit/). This classifier.py file should define a class named Classifier that has a fit() and a predict_proba() method. This workflow will run fit() on the training data and at testing time, run predict_proba() on the test data, using the trained model. If you are using a scikit-learn function, you can simply call the fit() and predict_proba() methods of the model you are using.

  • FeatureExtractor() - this workflow is designed for preprocessing data, for example converting non-numeric features into numeric, normalising data and creating new features using existing features. This workflow is designed to be used with the Classifier or Regressor workflow. It will import the file named feature_extractor.py from the submission directory (e.g. submissions/starting_kit/). This feature_extractor.py file should define a class named FeatureExtractor with a fit() and a transform() method. This workflow will run fit() on the features and target of the data and run transform() on the features of training data. Note that fit() takes both the features and target of the data as input to enable feature engineering strategies such as target encoding during training time. The output of this workflow is the preprocessed features of the data.

  • feature_extractor_regressor() - this workflow combines the FeatureExtrator() and Regressor() workflows such that data is first preprocessed with FeatureExtractor() and then Regressor() performs model training and prediction. Note that the fit() method of FeatureExtractor() is only performed on training data but not test data.

  • feature_extractor_classifier() - this workflow combines the FeatureExtractor() and Classifier() workflows such that data is first preprocessed with FeatureExtractor() and then Classifier() performs model training and prediciton. As above the fit() method of FeatureExtractor() is only performed on training data but not test data.

There are also a number of much more specific workflows, some of which were designed for one specific challenge but can be re-used for similar challenges:

  • ImageClassifier() - this workflow is for image classification tasks, particularly for cases when the dataset cannot be stored in memory. This workflow will import two files from the submissions folder; image_preprocessor.py and batch_classifier.py. image_preprocessor.py should define a function called transform() which preprocesses images. It should take an image as input and output an image. Optionally, this file can also define a function called transform_test(), which is only used to preprocess images at test time. If this is not defined, transform() will be used at train and test time. batch_classifier.py should define a class called BatchClassifier with the methods fit() and predict_prob(). fit() should fit a model to batches of images (you can define batch size). For an example you can take a look at the MNIST or Pollenating insects challenges.

  • SimplifiedImageClassifier() - this is a simplified version of the above workflow where there is no image preprocessing step and instead of training and test batches of images, fit() and predict_proba() is performed on one image at a time. For an example, take a look at the MNIST simplified and Pollenating insects challenges.

  • ObjectDetector() - this workflow is used for image object detection tasks. It workflow imports one, object_detector.py, from the submissions folder, which should define a class, ObjectDetector, with fit() and predict() methods. It was used in the Mars craters challenge and the Astronomy tutorial.

  • Clusterer() - this workflow was used for the High-energy physics tracking challenge which aimed to cluster particle hits. This workflow imports the file named clusterer.py from the submissions directory. This file should define a class called Clusterer with fit() and predict_single_event() methods. fit() takes the features and the cluster ID of each sample as arguments to train the clustering model. At testing time, the each sample is sent to predict_single_event() separately and the predicted cluster assignments are joined with the sample ID (the first column of the features data) and returned.

  • ElNino() - this workflow was used for the El Nino challenge which used temperature data over time to predict future temperatures. The workflow consists of the TimeSeriesFeatureExtractor() then Regressor() workflows.

  • GridFeatureExtractorClassifier() - this workflow was used in the California rainfall challenge. It consists of the GridFeatureExtractor() then Classifier() workflows. This workflow is similar to feature_extractor_classifier() except that GridFeatureExtractor() takes as input 3 dimensional spatial grid data.

  • DrugSpectra() - this workflow was used for the Drug spectra challenge. It implements both the feature_extractor_regressor() and feature_extractor_classifier() workflows to perform a classification task and a regression task on the same dataset. The submissions directory requires 4 files named; feature_extractor_clf.py, classifier.py, feature_extractor_reg.py and regressor.py.

If the appropriate workflow class does not exist in RAMP workflow, you can define your own workflow class within the problem.py file. If it is not too specific,We would also encourage you to add your class to RAMP workflow so others can use it in future. See Contributing.

The Titanic challenge employed the feature_extractor_classifier() workflow. This can be specified simply with:

workflow = rw.workflows.FeatureExtractorClassifier()
  1. Select score types Score types are metrics used to assess each submission. A large number of different score metrics are available. To use one or more existing score metrics, simply provide a list of the class names of the score you wish to use and assign this to a variable called score_types. For example, the Titanic challenge used 3 different score metrics:

    score_types = [
        rw.score_types.ROCAUC(name='auc'),
        rw.score_types.Accuracy(name='acc'),
        rw.score_types.NegativeLogLikelihood(name='nll'),
    ]
    

    If you select more than one score, all the score metrics will be calculated when you enter a submission to RAMP. You can select one score metric to be used as the official score, used to rank participants, or calculate a weighted combined score from the various score metrics. For example, the Drug spectra challenge used a weighted combination of ClassificationError and MARE (Mean Absolute Relative Error):

    score_type_1 = rw.score_types.ClassificationError(name='err', precision=3)
    score_type_2 = rw.score_types.MARE(name='mare', precision=3)
    score_types = [
        # The official score combines the two scores with weights 2/3 and 1/3.
        rw.score_types.Combined(
            name='combined', score_types=[score_type_1, score_type_2],
            weights=[2. / 3, 1. / 3], precision=3),
    ]
    

Note that the actual implementation was more complex as this challenge consisted of both a classification and regression task. For the purposes of this example, the extra complexity was ignored.

Again if the appropriate score metric class does not exist in RAMP workflow, you can define your own score metric class within the problem.py file. If it is not too specific, we would also encourage you to add your class to RAMP workflow so others can use it in future. See Contributing.

  1. Specify a cross-validation scheme Specify a way to split the ‘train’ data into training and validation sets. This should be done by defining a get_cv() function that takes the feature and target data as parameters and returns indicies that can be used to split the data. If you are using a function with a random element, e.g., StratifiedShuffleSplit() from scikit-learn, it is important to set the random seed. This ensures that the train and valuidation data will be the same for all participants.

    For example, the Titanic challenge used StratifiedShuffleSplit():

    def get_cv(X, y):
        cv = StratifiedShuffleSplit(n_splits=8, test_size=0.2, random_state=57)
        return cv.split(X, y)
    
  1. Provide the I/O methods The problem.py file needs to define a get_train_data() and a get_test_data() function that reads in the training and test data. These functions will be used to ‘get data’ both locally and on the RAMP sever. For example, this was implemented in the Titanic challenge using:

    _target_column_name = 'Survived'
    _ignore_column_names = ['PassengerId']
    
    def _read_data(path, f_name):
        data = pd.read_csv(os.path.join(path, 'data', f_name))
        y_array = data[_target_column_name].values
        X_df = data.drop([_target_column_name] + _ignore_column_names, axis=1)
        return X_df, y_array
    
    
    def get_train_data(path='.'):
        f_name = 'train.csv'
        return _read_data(path, f_name)
    
    
    def get_test_data(path='.'):
        f_name = 'test.csv'
        return _read_data(path, f_name)
    

    The _read_data() is not strictly required and is acting as a helper function in the code above.