新增混淆矩阵绘图功能

[Jasonsey]

绘图建议使用PyG2Plot

1. 单一标签混淆矩阵绘制
2. 多标签混淆矩阵绘制，标签识别规则如下：
   • AB -> AB:
     • A->A ✅
     • B->B ✅
   • AB -> AC:
     • A->A ✅
     • B->C ❌
   • AB -> CD:
     • $\underbrace{A->C}_{0.5}$ ❌
     • $\underbrace{A->D}_{0.5}$ ❌
     • $\underbrace{B->C}_{0.5}$ ❌
     • $\underbrace{B->D}_{0.5}$ ❌

[Jasonsey]

如果类别特别多（100+）PyG2Plot会卡顿，可以尝试 seaborn.heatmap，实例代码:

a = np.arange(4).reshape((2,2))
y_label = pd.Index(['a', 'b'], name='y')
x_label = pd.Index(['a', 'b'], name='x')
df = pd.DataFrame(a, index=y_label, columns=x_label)

sns.heatmap(df).get_figure().savefig('./data/output/submit/test.jpg')

[Jasonsey]

sns实现细节:

1. 需要安装对应的字体，如果系统中没有该字体，得想办法提示用户都某个网站上下载字体文件，并安装
2. 配置

import matplotlib.pyplot as plt
import seaborn as sns

sns.set(rc={'figure.figsize': (40, 40)})
sns.set_style('whitegrid', {'font.sans-serif': ['SimHei', 'Arial']})
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False

def get_true_pred_data(model, ds_test):
    y_true, y_pred = [], []
    for item in ds_test:
        y_pred.append(model.predict(item[0], batch_size=Config.batch_size))
        y_true.append(item[1])
    y_pred = (np.vstack(y_pred) > 0.5).astype(np.int64)
    y_true = np.vstack(y_true)
    np.save(Config.y_true_path, y_true)
    np.save(Config.y_pred_path, y_pred)
    return y_true, y_pred


def get_matrix(y_true, y_pred):
    matrix = np.zeros((len(Config.labels), len(Config.labels)))
    for _y_true, _y_pred in zip(y_true, y_pred):
        _y_true = np.argwhere(_y_true == 1).flatten()
        _y_pred = np.argwhere(_y_pred == 1).flatten()
        num = len(_y_true) * len(_y_pred)
        for _y_true_ in _y_true:
            for _y_pred_ in _y_pred:
                matrix[_y_true_, _y_pred_] += (1/num)
    matrix /= (np.sum(matrix, axis=1, keepdims=True) + 1.0e-5)
    return matrix


def get_plot_confusion_data(matrix):
    y_label = pd.Index(Config.labels, name='y_true')
    x_label = pd.Index(Config.labels, name='y_pred')
    df = pd.DataFrame(matrix, index=y_label, columns=x_label)
    return df


def plot_confusion_matrix(plot_data):
    heat_map = sns.heatmap(plot_data, vmax=1, cmap='GnBu', annot=True, fmt='.0%', annot_kws={'size': 6})
    heat_map.get_figure().savefig(Config.confusion_matrix_path, dpi=400)