Python learner.Learner类代码示例

def test_predict_on_subset_with_existing_model():
    """
    Test generating predictions on subset with existing model
    """
    # Create data files
    make_single_file_featureset_data()

    # train and save a model on the training file
    train_fs = NDJReader.for_path(join(_my_dir, 'train', 'train_single_file.jsonlines')).read()
    learner = Learner('RandomForestClassifier')
    learner.train(train_fs, grid_search=True, grid_objective="accuracy")
    model_filename = join(_my_dir, 'output', ('train_test_single_file_train_train_'
                                              'single_file.jsonlines_test_test_single'
                                              '_file_subset.jsonlines_RandomForestClassifier'
                                              '.model'))

    learner.save(model_filename)

    # Run experiment
    config_path = fill_in_config_paths_for_single_file(join(_my_dir, "configs",
                                                            "test_single_file_saved_subset"
                                                            ".template.cfg"),
                                                       join(_my_dir, 'train', 'train_single_file.jsonlines'),
                                                       join(_my_dir, 'test',
                                                            'test_single_file_subset.'
                                                            'jsonlines'))
    run_configuration(config_path, quiet=True, overwrite=False)

    # Check results
    with open(join(_my_dir, 'output', ('train_test_single_file_train_train_'
                                       'single_file.jsonlines_test_test_single'
                                       '_file_subset.jsonlines_RandomForestClassifier'
                                       '.results.json'))) as f:
        result_dict = json.load(f)[0]
    assert_almost_equal(result_dict['accuracy'], 0.7333333)

def check_specified_cv_folds(numeric_ids):
    make_cv_folds_data(numeric_ids)

    # test_cv_folds1.cfg has prespecified folds and should have ~50% accuracy
    # test_cv_folds2.cfg doesn't have prespecified folds and >95% accuracy
    for experiment_name, test_func, grid_size in [('test_cv_folds1',
                                                   lambda x: x < 0.6,
                                                   3),
                                                  ('test_cv_folds2',
                                                   lambda x: x > 0.95,
                                                   10)]:
        config_template_file = '{}.template.cfg'.format(experiment_name)
        config_template_path = os.path.join(_my_dir, 'configs',
                                            config_template_file)
        config_path = os.path.join(_my_dir,
                                   fill_in_config_paths(config_template_path))

        # Modify config file to change ids_to_floats depending on numeric_ids
        # setting
        with open(config_path, 'r+') as config_template_file:
            lines = config_template_file.readlines()
            config_template_file.seek(0)
            config_template_file.truncate()
            for line in lines:
                if line.startswith('ids_to_floats='):
                    if numeric_ids:
                        line = 'ids_to_floats=true\n'
                    else:
                        line = 'ids_to_floats=false\n'
                config_template_file.write(line)

        run_configuration(config_path, quiet=True)
        result_filename = ('{}_test_cv_folds_LogisticRegression.' +
                           'results').format(experiment_name)
        with open(os.path.join(_my_dir, 'output', result_filename)) as f:
            # check held out scores
            outstr = f.read()
            score = float(SCORE_OUTPUT_RE.search(outstr).groups()[-1])
            assert test_func(score)

            grid_score_matches = GRID_RE.findall(outstr)
            assert len(grid_score_matches) == grid_size
            for match_str in grid_score_matches:
                assert test_func(float(match_str))

    # try the same tests for just training (and specifying the folds for the
    # grid search)
    dirpath = os.path.join(_my_dir, 'train')
    suffix = '.jsonlines'
    featureset = ['test_cv_folds']
    examples = _load_featureset(dirpath, featureset, suffix, quiet=True)
    clf = Learner('LogisticRegression', probability=True)
    cv_folds = _load_cv_folds(os.path.join(_my_dir, 'train',
                                           'test_cv_folds.csv'))
    grid_search_score = clf.train(examples, grid_search_folds=cv_folds,
                                  grid_objective='accuracy', grid_jobs=1)
    assert grid_search_score < 0.6
    grid_search_score = clf.train(examples, grid_search_folds=5,
                                  grid_objective='accuracy', grid_jobs=1)
    assert grid_search_score > 0.95

def check_train_and_score_function(model_type):
    """
    Check that the _train_and_score() function works as expected
    """

    # create train and test data
    (train_fs,
     test_fs) = make_classification_data(num_examples=500,
                                         train_test_ratio=0.7,
                                         num_features=5,
                                         use_feature_hashing=False,
                                         non_negative=True)

    # call _train_and_score() on this data
    estimator_name = 'LogisticRegression' if model_type == 'classifier' else 'Ridge'
    metric = 'accuracy' if model_type == 'classifier' else 'pearson'
    learner1 = Learner(estimator_name)
    train_score1, test_score1 = _train_and_score(learner1, train_fs, test_fs, metric)

    # this should yield identical results when training another instance
    # of the same learner without grid search and shuffling and evaluating
    # that instance on the train and the test set
    learner2 = Learner(estimator_name)
    learner2.train(train_fs, grid_search=False, shuffle=False)
    train_score2 = learner2.evaluate(train_fs, output_metrics=[metric])[-1][metric]
    test_score2 = learner2.evaluate(test_fs, output_metrics=[metric])[-1][metric]

    eq_(train_score1, train_score2)
    eq_(test_score1, test_score2)

def check_scaling_features(use_feature_hashing=False, use_scaling=False):
    train_fs, test_fs = make_scaling_data(
        use_feature_hashing=use_feature_hashing)

    # create a Linear SVM with the value of scaling as specified
    feature_scaling = 'both' if use_scaling else 'none'
    learner = Learner('SGDClassifier', feature_scaling=feature_scaling,
                      pos_label_str=1)

    # train the learner on the training set and test on the testing set
    learner.train(train_fs)
    test_output = learner.evaluate(test_fs)
    fmeasures = [test_output[2][0]['F-measure'],
                 test_output[2][1]['F-measure']]

    # these are the expected values of the f-measures, sorted
    if not use_feature_hashing:
        expected_fmeasures = ([0.7979797979797979, 0.80198019801980192] if
                              not use_scaling else
                              [0.94883720930232551, 0.94054054054054048])
    else:
        expected_fmeasures = ([0.83962264150943389, 0.81914893617021278] if
                              not use_scaling else
                              [0.88038277511961716, 0.86910994764397898])

    assert_almost_equal(expected_fmeasures, fmeasures)

def check_scaling_features(use_feature_hashing=False, use_scaling=False):
    train_fs, test_fs = make_scaling_data(use_feature_hashing=use_feature_hashing)

    # create a Linear SVM with the value of scaling as specified
    feature_scaling = 'both' if use_scaling else 'none'
    learner = Learner('SGDClassifier', feature_scaling=feature_scaling,
                      pos_label_str=1)

    # train the learner on the training set and test on the testing set
    learner.train(train_fs)
    test_output = learner.evaluate(test_fs)
    fmeasures = [test_output[2][0]['F-measure'],
                 test_output[2][1]['F-measure']]

    # these are the expected values of the f-measures, sorted
    if not use_feature_hashing:
        expected_fmeasures = ([0.77319587628865982, 0.78640776699029125] if
                              not use_scaling else
                              [0.94930875576036866, 0.93989071038251359])
    else:
        expected_fmeasures = ([0.42774566473988435, 0.5638766519823788] if
                              not use_scaling else
                              [0.87323943661971837, 0.85561497326203206])

    assert_almost_equal(expected_fmeasures, fmeasures)

def check_tree_models(name,
                      use_feature_hashing=False,
                      use_rescaling=False):

    # create a FeatureSet object with the data we want to use
    if use_feature_hashing:
        train_fs, test_fs, _ = make_regression_data(num_examples=5000,
                                                    num_features=10,
                                                    use_feature_hashing=True,
                                                    feature_bins=5)
    else:
        train_fs, test_fs, _ = make_regression_data(num_examples=2000,
                                                    num_features=3)

    # create the learner
    if use_rescaling:
        name = 'Rescaled' + name
    learner = Learner(name)

    # train it with the training feature set we created
    # make sure to set the grid objective to pearson
    learner.train(train_fs, grid_objective='pearson')

    # make sure that the feature importances are as expected.
    if name.endswith('DecisionTreeRegressor'):
        expected_feature_importances = ([0.37331461,
                                         0.08572699,
                                         0.2543484,
                                         0.1841172,
                                         0.1024928] if use_feature_hashing else
                                        [0.08931994,
                                         0.15545093,
                                         0.75522913])
        expected_cor_range = [0.5, 0.6] if use_feature_hashing else [0.9, 1.0]
    else:
        if use_feature_hashing:
            expected_feature_importances = [0.40195655,
                                            0.06702161,
                                            0.25814858,
                                            0.18183947,
                                            0.09103379]
        else:
            expected_feature_importances = [0.07975691, 0.16122862, 0.75901447]
        expected_cor_range = [0.7, 0.8] if use_feature_hashing else [0.9, 1.0]

    feature_importances = learner.model.feature_importances_
    assert_allclose(feature_importances, expected_feature_importances,
                    rtol=1e-2)

    # now generate the predictions on the test FeatureSet
    predictions = learner.predict(test_fs)

    # now make sure that the predictions are close to
    # the actual test FeatureSet labels that we generated
    # using make_regression_data. To do this, we just
    # make sure that they are correlated with pearson > 0.95
    cor, _ = pearsonr(predictions, test_fs.labels)
    assert_greater(cor, expected_cor_range[0])
    assert_less(cor, expected_cor_range[1])

def check_tree_models(name,
                      use_feature_hashing=False,
                      use_rescaling=False):

    # create a FeatureSet object with the data we want to use
    if use_feature_hashing:
        train_fs, test_fs, _ = make_regression_data(num_examples=5000,
                                                    num_features=10,
                                                    use_feature_hashing=True,
                                                    feature_bins=5)
    else:
        train_fs, test_fs, _ = make_regression_data(num_examples=2000,
                                                    num_features=3)

    # create the learner
    if use_rescaling:
        name = 'Rescaled' + name
    learner = Learner(name)

    # train it with the training feature set we created
    # make sure to set the grid objective to pearson
    learner.train(train_fs, grid_objective='pearson')

    # make sure that the feature importances are as expected.
    if name.endswith('DecisionTreeRegressor'):
        expected_feature_importances = ([0.37483895,
                                         0.08816508,
                                         0.25379838,
                                         0.18337128,
                                         0.09982631] if use_feature_hashing else
                                        [0.08926899,
                                         0.15585068,
                                         0.75488033])
        expected_cor_range = [0.5, 0.6] if use_feature_hashing else [0.9, 1.0]
    else:
        expected_feature_importances = ([0.40195798,
                                         0.06702903,
                                         0.25816559,
                                         0.18185518,
                                         0.09099222] if use_feature_hashing else
                                        [0.07974267,
                                         0.16121895,
                                         0.75903838])
        expected_cor_range = [0.7, 0.8] if use_feature_hashing else [0.9, 1.0]

    feature_importances = learner.model.feature_importances_
    assert_allclose(feature_importances, expected_feature_importances,
                    atol=1e-2, rtol=0)

    # now generate the predictions on the test FeatureSet
    predictions = learner.predict(test_fs)

    # now make sure that the predictions are close to
    # the actual test FeatureSet labels that we generated
    # using make_regression_data. To do this, we just
    # make sure that they are correlated with pearson > 0.95
    cor, _ = pearsonr(predictions, test_fs.labels)
    assert_greater(cor, expected_cor_range[0])
    assert_less(cor, expected_cor_range[1])

def test_retrieve_cv_folds():
    """
    Test to make sure that the fold ids get returned correctly after cross-validation
    """

    # Setup
    learner = Learner('LogisticRegression')
    num_folds = 5
    cv_fs, custom_cv_folds = make_cv_folds_data(num_examples_per_fold=2, num_folds=num_folds)

    # Test 1: learner.cross_validate() makes the folds itself.
    expected_fold_ids = {'EXAMPLE_0': '0',
                         'EXAMPLE_1': '4',
                         'EXAMPLE_2': '3',
                         'EXAMPLE_3': '1',
                         'EXAMPLE_4': '2',
                         'EXAMPLE_5': '2',
                         'EXAMPLE_6': '1',
                         'EXAMPLE_7': '0',
                         'EXAMPLE_8': '4',
                         'EXAMPLE_9': '3'}
    _, _, _, skll_fold_ids = learner.cross_validate(cv_fs,
                                                    stratified=True,
                                                    cv_folds=num_folds,
                                                    grid_search=True,
                                                    grid_objective='f1_score_micro',
                                                    shuffle=False,
                                                    save_cv_folds=True)
    assert_equal(skll_fold_ids, expected_fold_ids)

    # Test 2: if we pass in custom fold ids, those are also preserved.
    _, _, _, skll_fold_ids = learner.cross_validate(cv_fs,
                                                    stratified=True,
                                                    cv_folds=custom_cv_folds,
                                                    grid_search=True,
                                                    grid_objective='f1_score_micro',
                                                    shuffle=False,
                                                    save_cv_folds=True)
    assert_equal(skll_fold_ids, custom_cv_folds)

    # Test 3: when learner.cross_validate() makes the folds but stratified=False
    # and grid_search=False, so that KFold is used.
    expected_fold_ids = {'EXAMPLE_0': '0',
                         'EXAMPLE_1': '0',
                         'EXAMPLE_2': '1',
                         'EXAMPLE_3': '1',
                         'EXAMPLE_4': '2',
                         'EXAMPLE_5': '2',
                         'EXAMPLE_6': '3',
                         'EXAMPLE_7': '3',
                         'EXAMPLE_8': '4',
                         'EXAMPLE_9': '4'}
    _, _, _, skll_fold_ids = learner.cross_validate(cv_fs,
                                                    stratified=False,
                                                    cv_folds=num_folds,
                                                    grid_search=False,
                                                    shuffle=False,
                                                    save_cv_folds=True)
    assert_equal(skll_fold_ids, custom_cv_folds)

def check_invalid_regr_grid_obj_func(learner_name, grid_objective_function):
    """
    Checks whether the grid objective function is valid for this regression
    learner
    """
    (train_fs, _, _) = make_regression_data()
    clf = Learner(learner_name)
    clf.train(train_fs, grid_objective=grid_objective_function)

def test_predict_dict_hasher():
    train_file = join(_my_dir, 'other', 'examples_train.jsonlines')
    test_file = join(_my_dir, 'other', 'examples_test.jsonlines')
    train_fs = NDJReader.for_path(train_file).read()
    test_fs = NDJReader.for_path(test_file, feature_hasher=True, num_features=3).read()
    learner = Learner('LogisticRegression')
    learner.train(train_fs, grid_search=False)
    _ = learner.predict(test_fs)

def check_ensemble_models(name,
                          use_feature_hashing=False,
                          use_rescaling=False):

    # create a FeatureSet object with the data we want to use
    if use_feature_hashing:
        train_fs, test_fs, _ = make_regression_data(num_examples=5000,
                                                    num_features=10,
                                                    use_feature_hashing=True,
                                                    feature_bins=5)
    else:
        train_fs, test_fs, _ = make_regression_data(num_examples=2000,
                                                    num_features=3)

    # create the learner
    if use_rescaling:
        name = 'Rescaled' + name
    learner = Learner(name)

    # train it with the training feature set we created
    # make sure to set the grid objective to pearson
    learner.train(train_fs, grid_objective='pearson')

    # make sure that the feature importances are as expected.
    if name.endswith('AdaBoostRegressor'):
        if use_feature_hashing:
            expected_feature_importances = [0.33718443,
                                            0.07810721,
                                            0.25621769,
                                            0.19489766,
                                            0.13359301]
        else:
            expected_feature_importances = [0.10266744, 0.18681777, 0.71051479]
    else:
        expected_feature_importances = ([0.204,
                                         0.172,
                                         0.178,
                                         0.212,
                                         0.234] if use_feature_hashing else
                                        [0.262,
                                         0.288,
                                         0.45])

    feature_importances = learner.model.feature_importances_
    assert_allclose(feature_importances, expected_feature_importances,
                    atol=1e-2, rtol=0)

    # now generate the predictions on the test FeatureSet
    predictions = learner.predict(test_fs)

    # now make sure that the predictions are close to
    # the actual test FeatureSet labels that we generated
    # using make_regression_data. To do this, we just
    # make sure that they are correlated with pearson > 0.95
    cor, _ = pearsonr(predictions, test_fs.labels)
    expected_cor_range = [0.7, 0.8] if use_feature_hashing else [0.9, 1.0]
    assert_greater(cor, expected_cor_range[0])
    assert_less(cor, expected_cor_range[1])

def test_predict_hasher_hasher_same_bins():
    train_file = join(_my_dir, 'other', 'examples_train.jsonlines')
    test_file = join(_my_dir, 'other', 'examples_test.jsonlines')
    train_fs = NDJReader.for_path(train_file, feature_hasher=True, num_features=3).read()
    test_fs = NDJReader.for_path(test_file, feature_hasher=True, num_features=3).read()
    learner = Learner('LogisticRegression')
    learner.train(train_fs, grid_search=False)
    predictions = learner.predict(test_fs)
    eq_(len(predictions), test_fs.features.shape[0])

def check_linear_models(name,
                        use_feature_hashing=False,
                        use_rescaling=False):

    # create a FeatureSet object with the data we want to use
    if use_feature_hashing:
        (train_fs,
         test_fs,
         weightdict) = make_regression_data(num_examples=5000,
                                            num_features=10,
                                            use_feature_hashing=True,
                                            feature_bins=5)
    else:
        train_fs, test_fs, weightdict = make_regression_data(num_examples=2000,
                                                             num_features=3)

    # create the learner
    if use_rescaling:
        name = 'Rescaled' + name
    learner = Learner(name)

    # train it with the training feature set we created
    # make sure to set the grid objective to pearson
    learner.train(train_fs, grid_search=True, grid_objective='pearson')

    # make sure that the weights are close to the weights
    # that we got from make_regression_data. Take the
    # ceiling before comparing since just comparing
    # the ceilings should be enough to make sure nothing
    # catastrophic happened. However, sometimes with
    # feature hashing, the ceiling is not exactly identical
    # so when that fails we want to check that the rounded
    # feature values are the same. One of those two equalities
    # _must_ be satisified.

    # get the weights for this trained model
    learned_weights = learner.model_params[0]

    for feature_name in learned_weights:
        learned_w_ceil = math.ceil(learned_weights[feature_name])
        given_w_ceil = math.ceil(weightdict[feature_name])
        learned_w_round = round(learned_weights[feature_name], 0)
        given_w_round = round(weightdict[feature_name], 0)
        ceil_equal = learned_w_ceil == given_w_ceil
        round_equal = learned_w_round == given_w_round
        either_equal = ceil_equal or round_equal
        assert either_equal

    # now generate the predictions on the test FeatureSet
    predictions = learner.predict(test_fs)

    # now make sure that the predictions are close to
    # the actual test FeatureSet labels that we generated
    # using make_regression_data. To do this, we just
    # make sure that they are correlated with pearson > 0.95
    cor, _ = pearsonr(predictions, test_fs.labels)
    assert_greater(cor, 0.95)

def check_ensemble_models(name,
                          use_feature_hashing=False,
                          use_rescaling=False):

    # create a FeatureSet object with the data we want to use
    if use_feature_hashing:
        train_fs, test_fs, _ = make_regression_data(num_examples=5000,
                                                    num_features=10,
                                                    use_feature_hashing=True,
                                                    feature_bins=5)
    else:
        train_fs, test_fs, _ = make_regression_data(num_examples=2000,
                                                    num_features=3)

    # create the learner
    if use_rescaling:
        name = 'Rescaled' + name
    learner = Learner(name)

    # train it with the training feature set we created
    # make sure to set the grid objective to pearson
    learner.train(train_fs, grid_search=True, grid_objective='pearson')

    # make sure that the feature importances are as expected.
    if name.endswith('AdaBoostRegressor'):
        if use_feature_hashing:
            expected_feature_importances = [0.749811,
                                            0.001373,
                                            0.23357,
                                            0.011691,
                                            0.003554]
        else:
            expected_feature_importances = [0.10266744, 0.18681777, 0.71051479]
    else:
        expected_feature_importances = ([0.735756,
                                         0.001034,
                                         0.242734,
                                         0.015836,
                                         0.00464] if use_feature_hashing else
                                        [0.082621,
                                         0.166652,
                                         0.750726])

    feature_importances = learner.model.feature_importances_
    assert_allclose(feature_importances, expected_feature_importances,
                    atol=1e-2, rtol=0)

    # now generate the predictions on the test FeatureSet
    predictions = learner.predict(test_fs)

    # now make sure that the predictions are close to
    # the actual test FeatureSet labels that we generated
    # using make_regression_data. To do this, we just
    # make sure that they are correlated with pearson > 0.95
    cor, _ = pearsonr(predictions, test_fs.labels)
    assert_greater(cor, 0.95)

def check_tree_models(name,
                      use_feature_hashing=False,
                      use_rescaling=False):

    # create a FeatureSet object with the data we want to use
    if use_feature_hashing:
        train_fs, test_fs, _ = make_regression_data(num_examples=5000,
                                                    num_features=10,
                                                    use_feature_hashing=True,
                                                    feature_bins=5)
    else:
        train_fs, test_fs, _ = make_regression_data(num_examples=2000,
                                                    num_features=3)

    # create the learner
    if use_rescaling:
        name = 'Rescaled' + name
    learner = Learner(name)

    # train it with the training feature set we created
    # make sure to set the grid objective to pearson
    learner.train(train_fs, grid_search=True, grid_objective='pearson')

    # make sure that the feature importances are as expected.
    if name.endswith('DecisionTreeRegressor'):
        expected_feature_importances = ([0.730811,
                                         0.001834,
                                         0.247603,
                                         0.015241,
                                         0.004511] if use_feature_hashing else
                                        [0.08926899,
                                         0.15585068,
                                         0.75488033])
    else:
        expected_feature_importances = ([0.733654,
                                         0.002528,
                                         0.245527,
                                         0.013664,
                                         0.004627] if use_feature_hashing else
                                        [0.07974267,
                                         0.16121895,
                                         0.75903838])

    feature_importances = learner.model.feature_importances_
    assert_allclose(feature_importances, expected_feature_importances,
                    atol=1e-2, rtol=0)

    # now generate the predictions on the test FeatureSet
    predictions = learner.predict(test_fs)

    # now make sure that the predictions are close to
    # the actual test FeatureSet labels that we generated
    # using make_regression_data. To do this, we just
    # make sure that they are correlated with pearson > 0.95
    cor, _ = pearsonr(predictions, test_fs.labels)
    assert_greater(cor, 0.95)