Ejemplo n.º 1
0
def main():
    p = optparse.OptionParser()
    p.add_option('--model', '-m', default = 'model', type = str, help = 'model filename prefix')
    p.add_option('--load', '-L', default = False, action = 'store_true', help = 'load model from file')
    p.add_option('--features', '-f', default = 'features.txt', type = str, help = 'feature filename')
    p.add_option('--data', '-d', default = 'data.csv', type = str, help = 'marked data filename')
    p.add_option('--verbose', '-v', default = False, action = 'store_true', help = 'verbosity flag')
    p.add_option('--thresh', '-T', default = 0.5, type = float, help = 'probability threshold to classify True')
    p.add_option('--n_estimators', '-n', default = 100, type = int, help = 'number of random forest estimators')
    p.add_option('--test_fraction', '-t', default = 0.25, type = float, help = 'fraction of data to use for testing')
    p.add_option('--seed', '-s', default = None, type = int, help = 'random seed')
    p.add_option('--jobs', '-j', default = -1, type = int, help = 'number of jobs (-1 if maximum)')
    p.add_option('--probs', '-p', default = None, type = str, help = 'filename for output probabilities')
    opts, args = p.parse_args()

    model_filename = opts.model + '%s.pickle' % ('' if opts.seed is None else str(opts.seed))
    probs_filename = ('predicted_probs%s.dat' % ('' if opts.seed is None else str(opts.seed))) if opts.probs is None else opts.probs

    np.random.seed(opts.seed)

    if opts.verbose:
        print("\nReading marked data from %s..." % opts.data)

    # establish data frame
    df = pd.read_csv(opts.data)
    n_lines = len(df)

    # choose test set as random test_fraction of data, leaving the remainder for training
    n_test = int(opts.test_fraction * n_lines)
    if opts.verbose:
        print("Read %d lines of data -> %d lines (training), %d lines (test)" % (n_lines, n_lines - n_test, n_test))
    test_subset = np.random.permutation(range(n_lines))[:n_test]
    is_train = np.ones(n_lines, dtype = bool)
    for i in test_subset:
        is_train[i] = False

    # establish training and test sets
    train, test = df[is_train], df[~is_train]

    if opts.load:
        rfc = pickle.load(open(model_filename, 'rb'))
        if opts.verbose:
            print("\nLoaded model from '%s'.\n" % model_filename)

    else:
        # set the random forest instance
        rfc = RandomForestClassifier(n_estimators = opts.n_estimators, n_jobs = opts.jobs)
        # set list of features (all the uncommented features above dotted line in feature file; leading/trailing whitespace is stripped
        with open(opts.features, 'r') as f:
            lines = f.readlines()
        line_starts_with_dash = [(line[0] == '-') for line in lines]
        assert (line_starts_with_dash.count(True) == 1), "Feature file must have a single dashed line separating input/output features."
        dashed_line_index = line_starts_with_dash.index(True)
        rfc.input_features = []
        for i in range(dashed_line_index):
            feature = lines[i].partition('#')[0].strip()
            if (len(feature) > 0):
                rfc.input_features.append(feature)
        output_features = []
        for i in range(dashed_line_index + 1, len(lines)):
            feature = lines[i].partition('#')[0].strip()
            if (len(feature) > 0):
                output_features.append(feature)
        assert (len(output_features) == 1), "Feature file must have exactly one output feature."
        rfc.output_feature = output_features[0]

    num_features = len(rfc.input_features)
    assert (num_features > 0), "Feature file must have at least one input feature."
    X = train[rfc.input_features]
    y = train[rfc.output_feature]

    if (not opts.load):
        # train the forest
        if opts.verbose:
            print("\nTraining %d random forests..." % opts.n_estimators)
        rfc.fit(X, y)
        # save off the model
        pickle.dump(rfc, open(model_filename, 'wb'))
        if opts.verbose:
            print("\nSaved model to '%s'.\n" % model_filename)

    # make predictions on the test data
    probs = rfc.predict_proba(test[rfc.input_features])[:, 1]
    probs_series = pd.Series(probs)
    probs_series.to_csv(probs_filename, index = False)
    test_preds = (probs >= opts.thresh)
    conf_df = pd.crosstab(test[rfc.output_feature], test_preds, rownames = ['actual'], colnames = ['predicted'])
    conf_mat = np.asarray(conf_df)
    class_report = classification_report(test[rfc.output_feature], test_preds)
    print("\nConfusion Matrix")
    print(conf_df)
    print("\nClassification Report")
    print(class_report)
    accuracy = (conf_mat[0, 0] + conf_mat[1, 1]) / float(np.sum(conf_mat))
    print("Accuracy = %.3f%%" % (100. * accuracy))
    print("\nFeature Importances")
    triples = [(i, rfc.input_features[i], rfc.feature_importances_[i]) for i in range(num_features)]
    triples.sort(key = lambda pair : pair[2], reverse = True)
    indices, features, importances = zip(*triples)
    for i in range(num_features):
        print("%17s   %3d.%03d%%" % (features[i], int(100. * importances[i]), round(1000 * (100. * importances[i] - int(100. * importances[i])))))
    stds = np.std([tree.feature_importances_ for tree in rfc.estimators_], axis = 0)
    plt.figure()
    plt.title("Feature importances")
    plt.bar(range(X.shape[1]), importances, color = 'r', yerr = [stds[i] for i in indices], align = 'center')
    plt.xticks(range(X.shape[1]), features, rotation = 'vertical')
    plt.xlim([-1, X.shape[1]])
    fig = plt.gcf()
    fig.subplots_adjust(bottom = 0.25)
    plt.savefig('feature_importances%s.png' % ('' if opts.seed is None else str(opts.seed)))
Ejemplo n.º 2
0
def main():
    p = optparse.OptionParser()
    p.add_option('--load', '-L', default = False, action = 'store_true', help = 'load model from file')
    p.add_option('--features', '-f', default = 'features.txt', type = str, help = 'feature filename')
    p.add_option('--verbose', '-v', default = False, action = 'store_true', help = 'verbosity flag')
    p.add_option('--thresh', '-T', default = 0.5, type = float, help = 'probability threshold to classify True')
    p.add_option('--n_estimators', '-n', default = 100, type = int, help = 'number of random forest estimators')
    p.add_option('--seed', '-s', default = None, type = int, help = 'random seed')
    p.add_option('--jobs', '-j', default = -1, type = int, help = 'number of jobs (-1 if maximum)')
    opts, args = p.parse_args()

    model_filename = 'model%s.pickle' % ('' if opts.seed is None else str(opts.seed))

    np.random.seed(opts.seed)

    if opts.verbose:
        print("\nReading data set...")

    train = pd.read_csv('yoochoose/data/training_session_features.csv').append(pd.read_csv('yoochoose/data/dev_session_features.csv'))
    test = pd.read_csv('yoochoose/data/test_session_features.csv')

    if opts.load:
        rfc = pickle.load(open(model_filename, 'rb'))
        if opts.verbose:
            print("\nLoaded model from '%s'.\n" % model_filename)
    else:
        # set the random forest instance
        rfc = RandomForestClassifier(n_estimators = opts.n_estimators, n_jobs = opts.jobs)
        # set list of features (all the uncommented features above dotted line in feature file; leading/trailing whitespace is stripped
        with open(opts.features, 'r') as f:
            lines = f.readlines()
        line_starts_with_dash = [(line[0] == '-') for line in lines]
        assert (line_starts_with_dash.count(True) == 1), "Feature file must have a single dashed line separating input/output features."
        dashed_line_index = line_starts_with_dash.index(True)
        rfc.input_features = []
        for i in range(dashed_line_index):
            feature = lines[i].partition('#')[0].strip()
            if (len(feature) > 0):
                rfc.input_features.append(feature)
        output_features = []
        for i in range(dashed_line_index + 1, len(lines)):
            feature = lines[i].partition('#')[0].strip()
            if (len(feature) > 0):
                output_features.append(feature)
        assert (len(output_features) == 1), "Feature file must have exactly one output feature."
        rfc.output_feature = output_features[0]

    num_features = len(rfc.input_features)
    assert (num_features > 0), "Feature file must have at least one input feature."
    X = train[rfc.input_features]
    y = train[rfc.output_feature]

    if (not opts.load):
        # train the forest
        if opts.verbose:
            print("\nTraining %d random forests..." % opts.n_estimators)
        rfc.fit(X, y)
        # save off the model
        pickle.dump(rfc, open(model_filename, 'wb'))
        if opts.verbose:
            print("\nSaved model to '%s'.\n" % model_filename)

    # make predictions on the test data
    probs = rfc.predict_proba(test[rfc.input_features])[:, 1]
    probs_series = pd.Series(probs)
    probs_series.to_csv('test_probs%s' % ('' if opts.seed is None else str(opts.seed)), index = False)
    test_preds = (probs >= opts.thresh)
    conf_df = pd.crosstab(test[rfc.output_feature], test_preds, rownames = ['actual'], colnames = ['predicted'])
    conf_mat = np.asarray(conf_df)
    class_report = classification_report(test[rfc.output_feature], test_preds)
    s = "\nConfusion Matrix\n"
    s += str(conf_df) + '\n'
    s += "\nClassification Report\n"
    s += class_report + '\n'
    accuracy = (conf_mat[0, 0] + conf_mat[1, 1]) / float(np.sum(conf_mat))
    s += "Accuracy = %.3f%%\n" % (100. * accuracy)
    s += "\nFeature Importances\n"
    triples = [(i, rfc.input_features[i], rfc.feature_importances_[i]) for i in range(num_features)]
    triples.sort(key = lambda pair : pair[2], reverse = True)
    indices, features, importances = zip(*triples)
    for i in range(num_features):
        s += "%17s   %3d.%03d%%\n" % (features[i], int(100. * importances[i]), round(1000 * (100. * importances[i] - int(100. * importances[i]))))
    with open('test_report%s' % ('' if opts.seed is None else str(opts.seed)), 'w') as f:
        f.write(s)