_hyperplot.py

"""*Matplotlib* wrapper tools for optimization of
hyperparameters results display and analysis.

"""
import json
import math
import os
from os import path

import numpy as np


def _get_results(exp):
    report_path = path.join(exp, "results")
    results = []
    for file in os.listdir(report_path):
        if path.isfile(path.join(report_path, file)):
            with open(path.join(report_path, file), "r") as f:
                results.append(json.load(f))
    return results


def _outliers_idx(values, max_deviation):
    mean = values.mean()

    dist = abs(values - mean)
    corrected = dist < mean + max_deviation
    return corrected


def _logscale_plot(ax, xrange, yrange, base=10):

    from matplotlib import ticker

    if xrange is not None:
        ax.xaxis.set_minor_formatter(ticker.LogFormatter())
        ax.xaxis.set_major_formatter(ticker.LogFormatter())
        ax.set_xscale("log", base=base)
        ax.set_xlim([np.min(xrange), np.max(xrange)])
    if yrange is not None:
        ax.yaxis.set_minor_formatter(ticker.LogFormatter())
        ax.yaxis.set_major_formatter(ticker.LogFormatter())
        ax.set_yscale("log", base=base)
        ax.set_ylim(
            [yrange.min() - 0.1 * yrange.min(), yrange.max() + 0.1 * yrange.min()]
        )


def _scale(x):
    return (x - x.min()) / (x.ptp())


def _cross_parameter_plot(
    ax, values, scores, loss, smaxs, cmaxs, lmaxs, p1, p2, log1, log2, cat1, cat2
):

    X = values[p2].copy()
    Y = values[p1].copy()

    to_log = []
    if log1 and not cat2:
        to_log.append(X)
    else:
        to_log.append(None)
        if cat2:
            ax.margins(x=0.05)
    if log1 and not cat1:
        to_log.append(Y)
    else:
        to_log.append(None)

    _logscale_plot(ax, *to_log)
    # if log2 and not cat2:
    #     logscale_plot(ax, None, Y)

    ax.tick_params(axis="both", which="both")
    ax.tick_params(axis="both", labelsize="xx-small")
    ax.grid(True, which="both", ls="-", alpha=0.75)

    ax.set_xlabel(p2)
    ax.set_ylabel(p1)

    sc_l = ax.scatter(
        X[lmaxs], Y[lmaxs], scores[lmaxs] * 100, c=loss[lmaxs], cmap="inferno"
    )
    sc_s = ax.scatter(X[smaxs], Y[smaxs], scores[smaxs] * 100, c=cmaxs, cmap="YlGn")
    sc_m = ax.scatter(X[~(lmaxs)], Y[~(lmaxs)], scores[~(lmaxs)] * 100, color="red")

    return sc_l, sc_s, sc_m


def _loss_plot(
    ax,
    values,
    scores,
    loss,
    smaxs,
    cmaxs,
    lmaxs,
    p,
    log,
    categorical,
    legend,
    loss_behaviour,
):

    X = values[p].copy()

    if log and not (categorical):
        _logscale_plot(ax, X, loss)
    else:
        _logscale_plot(ax, None, loss)
        if categorical:
            ax.margins(x=0.05)

    ax.set_xlabel(p)
    ax.set_ylabel("loss")

    ax.tick_params(axis="both", which="both")
    ax.tick_params(axis="both", labelsize="xx-small")
    ax.grid(True, which="both", ls="-", alpha=0.75)

    sc_l = ax.scatter(X[lmaxs], loss[lmaxs], scores[lmaxs] * 100, color="orange")
    sc_s = ax.scatter(X[smaxs], loss[smaxs], scores[smaxs] * 100, c=cmaxs, cmap="YlGn")
    if loss_behaviour == "min":
        sc_m = ax.scatter(
            X[~(lmaxs)],
            [loss.min()] * np.sum(~lmaxs),
            scores[~(lmaxs)] * 100,
            color="red",
            label="Loss min.",
        )
    else:
        sc_m = ax.scatter(
            X[~(lmaxs)],
            [loss.max()] * np.sum(~lmaxs),
            scores[~(lmaxs)] * 100,
            color="red",
            label="Score max.",
        )

    if legend:
        ax.legend()

    return sc_l, sc_s, sc_m


def _parameter_violin(ax, values, scores, loss, smaxs, cmaxs, p, log, legend):

    import matplotlib.pyplot as plt

    y = values[p].copy()[smaxs]
    all_y = values[p].copy()

    if log:
        y = np.log10(y)
        all_y = np.log10(all_y)

    ax.get_yaxis().set_ticks([])
    ax.tick_params(axis="x", which="both")

    ax.xaxis.set_major_locator(plt.MultipleLocator(1))

    def format_func(value, tick_number):
        return "$10^{" + str(int(np.floor(value))) + "}$"

    ax.xaxis.set_major_formatter(plt.FuncFormatter(format_func))

    ax.set_xlabel(p)
    ax.grid(True, which="both", ls="-", alpha=0.75)
    ax.set_xlim([np.min(all_y), np.max(all_y)])

    violin = ax.violinplot(y, vert=False, showmeans=False, showextrema=False)

    for pc in violin["bodies"]:
        pc.set_facecolor("forestgreen")
        pc.set_edgecolor("white")

    quartile1, medians, quartile3 = np.percentile(y, [25, 50, 75])

    ax.scatter(medians, 1, marker="o", color="orange", s=30, zorder=4, label="Median")
    ax.hlines(
        1, quartile1, quartile3, color="gray", linestyle="-", lw=4, label="Q25/Q75"
    )
    ax.vlines(y.mean(), 0.5, 1.5, color="blue", label="Mean")

    ax.scatter(
        np.log10(values[p][scores.argmax()]),
        1,
        color="red",
        zorder=5,
        label="Best score",
    )
    ax.scatter(y, np.ones_like(y), c=cmaxs, cmap="YlGn", alpha=0.5, zorder=3)

    if legend:
        ax.legend(loc=2)


def _parameter_bar(ax, values, scores, loss, smaxs, cmaxs, p, categories):

    y = values[p].copy()[smaxs]

    ax.set_xlabel(p)
    ax.grid(True, which="both", ls="-", alpha=0.75)

    heights = []
    for p in categories:
        heights.append(y.tolist().count(p))

    ax.bar(x=categories, height=heights, color="forestgreen", alpha=0.3)


def plot_hyperopt_report(
    exp,
    params,
    metric="loss",
    loss_metric="loss",
    loss_behaviour="min",
    not_log=None,
    categorical=None,
    max_deviation=None,
    title=None,
):
    """Cross parameter scatter plot of hyperopt trials.

    Note
    ----
        Installation of Matplotlib and Seaborn packages
        is required to use this tool.

    Parameters
    ----------
        exp : str or Path
            Report directory storing hyperopt trials results.

        params : list
            Parameters to plot.

        metric : str, optional
            Metric to use as performance measure,
            stored in the hyperopt trials results dictionaries.
            May be different from loss metric. By default,
            'loss' is used as performance metric.

        loss_metric : str, optional
            Metric to use as an error measure,
            stored in the hyperopt trials results dictionaries.
            May be different from the default `loss` parameter.

        loss_behaviour : {'min', 'max'}, optional
            How to interpret metric used as main loss in the plot.
            If loss need to be minimized, choose 'min'. If loss
            need to be maximized, choose 'max'. In most cases,
            the loss is an error function that needs to be
            minimized. By default, 'min'.

        not_log : list, optional
            List of parameters to plot with a linear scale. By default,
            all scales are logarithmic.

        categorical : list, optional
            List of parameters to interpret as categorical or
            discrete valued.

        max_deviation : float, optional
            Maximum standard deviation expected from the loss mean.
            Useful to remove extreme outliers that may create odd plots.
            By default, all values are kept and plotted.

        title : str, optional
            Optional title for the figure.

    Returns:
        matplotlib.pyplot.figure
            Matplotlib figure object.

    """
    import matplotlib.pyplot as plt
    import seaborn as sns

    sns.set(context="paper", style="darkgrid", font_scale=1.5)
    N = len(params)
    not_log = not_log or []

    results = _get_results(exp)

    loss = np.array([r["returned_dict"][loss_metric] for r in results])
    scores = np.array([r["returned_dict"][metric] for r in results])

    if max_deviation is not None:
        not_outliers = _outliers_idx(loss, max_deviation)
        loss = loss[not_outliers]
        scores = scores[not_outliers]
        values = {
            p: np.array([r["current_params"][p] for r in results])[not_outliers]
            for p in params
        }
    else:
        values = {
            p: np.array([r["current_params"][p] for r in results]) for p in params
        }

    categorical = categorical or []

    for p in categorical:
        values[p] = values[p].astype(str)

    # Sorting for categorical plotting
    all_categorical = [val for p, val in values.items() if p in categorical]
    all_numerical = [val for p, val in values.items() if p not in categorical]

    sorted_idx = np.lexsort((loss, scores, *all_numerical, *all_categorical))

    loss = np.array([loss[i] for i in sorted_idx])
    scores = np.array([scores[i] for i in sorted_idx])

    for p, val in values.items():
        values[p] = np.array([val[i] for i in sorted_idx])

    scores = _scale(scores)

    # loss and f1 values

    if loss_behaviour == "min":
        lmaxs = loss > loss.min()
    else:
        lmaxs = loss < loss.max()

    percent = math.ceil(len(scores) * 0.05)
    smaxs = scores.argsort()[-percent:][::-1]
    cmaxs = _scale(scores[smaxs])

    # gridspecs

    fig = plt.figure(figsize=(15, 19), constrained_layout=True)
    gs = fig.add_gridspec(2, 1, height_ratios=[2 / 30, 28 / 30])
    fig.suptitle(f"Hyperopt trials summary - {title}", size=15)

    gs0 = gs[0].subgridspec(1, 3)
    gs1 = gs[1].subgridspec(N + 1, N)

    lbar_ax = fig.add_subplot(gs0[0, 0])
    fbar_ax = fig.add_subplot(gs0[0, 1])
    rad_ax = fig.add_subplot(gs0[0, 2])
    rad_ax.axis("off")

    # plot
    axes = []
    for i, p1 in enumerate(params):
        for j, p2 in enumerate(params):
            ax = fig.add_subplot(gs1[i, j])
            axes.append(ax)
            if p1 == p2:
                sc_l, sc_s, sc_m = _loss_plot(
                    ax,
                    values,
                    scores,
                    loss,
                    smaxs,
                    cmaxs,
                    lmaxs,
                    p2,
                    not (p2 in not_log),
                    p2 in categorical,
                    (i == 0 and j == 0),
                    loss_behaviour,
                )
            else:
                sc_l, sc_s, sc_m = _cross_parameter_plot(
                    ax,
                    values,
                    scores,
                    loss,
                    smaxs,
                    cmaxs,
                    lmaxs,
                    p1,
                    p2,
                    not (p1 in not_log),
                    not (p2 in not_log),
                    p1 in categorical,
                    p2 in categorical,
                )

    # legends

    handles, labels = sc_l.legend_elements(prop="sizes")
    legend = rad_ax.legend(
        handles,
        labels,
        loc="center left",
        title=f"Normalized {metric} (%)",
        mode="expand",
        ncol=len(labels) // 2 + 1,
    )

    l_cbar = fig.colorbar(sc_l, cax=lbar_ax, ax=axes, orientation="horizontal")
    _ = l_cbar.ax.set_title("Loss value")

    f_cbar = fig.colorbar(
        sc_s, cax=fbar_ax, ax=axes, orientation="horizontal", ticks=[0, 0.5, 1]
    )
    _ = f_cbar.ax.set_title(f"{metric} best population")
    _ = f_cbar.ax.set_xticklabels(["5% best", "2.5% best", "Best"])

    # violinplots

    legend = True
    for i, p in enumerate(params):
        ax = fig.add_subplot(gs1[-1, i])
        if p in categorical:
            _parameter_bar(
                ax, values, scores, loss, smaxs, cmaxs, p, sorted(list(set(values[p])))
            )
        else:
            _parameter_violin(
                ax, values, scores, loss, smaxs, cmaxs, p, not (p in not_log), legend
            )
            legend = False
        if legend:
            ax.set_ylabel(f"5% best {metric}\nparameter distribution")
    return fig