Source code for skvisualizations.validation

"""
Scikit-learn-compatible visualizations for model validation.

@author: David Diaz Vico
@license: MIT
"""

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from sklearn.decomposition import PCA

sns.set(style="white", palette="muted", color_codes=True)




def classifier_scatter(X, y, fname, pca_n_components=2, **kwargs):
    """Classifier scatter.

    Classifier scatter plot.

    Parameters
    ----------
    X: array-like, shape (n_samples, features_shape)
       The transformed data.
    y: numpy array of shape [n_samples]
       Target values.
    fname: str or file-like object
           https://matplotlib.org/api/_as_gen/matplotlib.pyplot.savefig.html
    pca_n_components: integer, default=2
                      Dimension of the PCA projection of X.
    **kwargs: optional savefig named args

    Returns
    -------
    None.
    """
    names = list(range(X.shape[1]))
    if pca_n_components is not None:
        names = list(range(pca_n_components))
        X = PCA(n_components=pca_n_components).fit_transform(X)
    names.append("class")
    data = pd.DataFrame(data=np.append(X, np.reshape(y, (len(y), 1)), axis=1), columns=names)
    sns.set()
    sns.pairplot(data, hue="class", x_vars=names[:-1], y_vars=names[:-1])
    plt.savefig(fname, **kwargs)





def regressor_scatter(X, y, preds, fname, **kwargs):
    """Regressor scatter.

    Regressor scatter plot.

    Parameters
    ----------
    X: array-like, shape (n_samples, features_shape)
       The transformed data.
    y: numpy array of shape [n_samples]
       Target values.
    preds: numpy array of shape [n_samples]
           Predicted values.
    fname: str or file-like object
           https://matplotlib.org/api/_as_gen/matplotlib.pyplot.savefig.html
    **kwargs: optional savefig named args

    Returns
    -------
    None.
    """
    X = PCA(n_components=1).fit_transform(X)
    names = ("X", "y", "source")
    data_y = np.append(X, np.reshape(y, (len(y), 1)), axis=1)
    data_y = np.append(data_y, np.reshape([0] * len(y), (len(y), 1)), axis=1)
    data_preds = np.append(X, np.reshape(preds, (len(preds), 1)), axis=1)
    data_preds = np.append(data_preds, np.reshape([1] * len(preds), (len(preds), 1)), axis=1)
    data = pd.DataFrame(data=np.append(data_y, data_preds, axis=0), columns=names)
    sns.set()
    sns.lmplot(x="X", y="y", hue="source", data=data)
    plt.savefig(fname, **kwargs)





def metaparameter_plot(search, param, fname, score="score", log_scale=True, **kwargs):
    """Metaparameter plot.

    Train and test metric plotted along a meta-parameter search space.

    Parameters
    ----------
    search: search object
            Fitted sklearn search object.
    param: string
           Name of the meta-parameter.
    fname: str or file-like object
           https://matplotlib.org/api/_as_gen/matplotlib.pyplot.savefig.html
    score: string
           Name of the metric
    log_scale: boolean, default=True
               Wether to use a logarithmic scale.
    **kwargs: optional savefig named args

    Returns
    -------
    None.
    """
    param_range = search.cv_results_["param_" + param].data.astype("float32")
    train_mean = search.cv_results_["mean_train_" + score]
    train_std = search.cv_results_["std_train_" + score]
    test_mean = search.cv_results_["mean_test_" + score]
    test_std = search.cv_results_["std_test_" + score]
    plt.figure()
    if log_scale:
        plt.xscale("log")
    plt.xlabel(param)
    plt.ylabel(score)
    plt.plot(param_range, train_mean, "o", label="Train", color="b")
    plt.fill_between(param_range, train_mean - train_std, train_mean + train_std, alpha=0.2, color="b")
    plt.plot(param_range, test_mean, "o", label="Test", color="g")
    plt.fill_between(param_range, test_mean - test_std, test_mean + test_std, alpha=0.2, color="g")
    plt.plot(param_range[search.best_index_], test_mean[search.best_index_], "o", label="Best", color="r")
    plt.plot(param_range[search.best_index_], train_mean[search.best_index_], "o", color="r")
    plt.axvline(x=param_range[search.best_index_], color="r")
    plt.legend(loc="best")
    plt.savefig(fname, **kwargs)





def keras_history_plot(history, fname, **kwargs):
    """Keras history plot.

    Train loss plotted for each training epoch.

    Parameters
    ----------
    history: history object
             Keras history object.
    fname: str or file-like object
           https://matplotlib.org/api/_as_gen/matplotlib.pyplot.savefig.html
    **kwargs: optional savefig named args

    Returns
    -------
    None.
    """
    plt.figure()
    plt.xlabel("Epoch")
    plt.ylabel("Loss")
    for k, v in history.items():
        plt.plot(v, label=k)
    plt.legend(loc="best")
    plt.savefig(fname, **kwargs)