plot.py

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd


def plot_2d_space(
    model, one_hot_encoder, x, y, n=None, reverse=False,
    save_to=None, dpi=256,
):
    """
    For models trained on equations ala 'a + b', this will plot a scatter plot
    with correct examples in green and incorrect ones in red.
    """
    if n is None:
        n = len(x)

    order = -1 if reverse else 1

    predictions = model.predict(x[:n])

    correct_coords = []
    incorrect_coords = []

    for i, prediction in enumerate(predictions):
        target = y[i]

        equation_string = one_hot_to_string(x[i])[::order]
        prediction_string = one_hot_to_string(prediction)[::order]
        target_string = one_hot_to_string(target)[::order]

        equation_plus_index = equation_string.index('+')
        n1 = int(equation_string[:equation_plus_index-1])
        n2 = int(equation_string[equation_plus_index+1:-1])

        if prediction_string == target_string:
            # Correct
            correct_coords.append((n1, n2,))
        else:
            incorrect_coords.append((n1, n2,))

    # Create plot
    fig = plt.figure()
    ax = fig.add_subplot(1, 1, 1, axisbg="1.0")

    correct_coords = np.array(correct_coords)
    ax.scatter(
        correct_coords[:, 0],
        correct_coords[:, 1],
        alpha=0.33, c='green', edgecolors='none', s=20, label='correct'
    )

    incorrect_coords = np.array(incorrect_coords)
    ax.scatter(
        incorrect_coords[:, 0],
        incorrect_coords[:, 1],
        alpha=0.33, c='red', edgecolors='none', s=20, label='incorrect'
    )

    if save_to:
        plt.savefig(save_to, dpi=dpi)
    else:
        plt.show()

    plt.close()


def plot_error_histogram(
    model, one_hot_encoder, x, y, n=None, max_d=10, reverse=False
):
    """
    For models trained on equations ala 'a + b', this will plot a histogram of
    error differences.
    """
    if n is None:
        n = len(x)

    order = -1 if reverse else 1

    predictions = model.predict(x[:n])

    differences = [0] * max_d

    for i, prediction in enumerate(predictions):
        target = y[i]

        prediction_string = one_hot_to_string(prediction)[::order]
        target_string = one_hot_to_string(target)[::order]

        prediction_int = int(prediction_string.strip(' \x00'))
        target_int = int(target_string.strip(' \x00'))

        difference = abs(target_int - prediction_int)

        if difference >= max_d:
            difference = max_d - 1

        differences[difference] += 1

    # Create plot
    fig = plt.figure()

    n, bins, patches = plt.hist(differences, normed=1, facecolor='green')

    plt.show()
    plt.close()


def plot_training_log(
    log_file_name='log.csv', metric='acc', save_to=None, dpi=256,
):
    """
    Plot a training log csv generated by CSVLogger. Will display the plot,
    or if you pass a save_to filename, render to a file.
    """
    log = pd.read_csv(log_file_name)

    if metric == 'acc':
        col_train = log.columns[1]
        col_test = log.columns[3]
        title = 'Accuracy ({})'.format(col_train)

    elif metric == 'loss':
        col_train = log.columns[2]
        col_test = log.columns[4]
        title = 'Loss'

    else:
        raise ValueError('Metric not s')

    plt.plot(log[col_train])
    plt.plot(log[col_test])
    plt.title(title)
    plt.ylabel(metric)
    plt.xlabel('epoch')
    plt.legend(['train', 'test'], loc='lower right')

    if save_to:
        plt.savefig(save_to, dpi=dpi)
    else:
        plt.show()

    plt.close()