import seaborn as sns
from matplotlib import pyplot as plt
from scipy.special import softmax
from sklearn.metrics import accuracy_score, confusion_matrix, precision_score, recall_score, f1_score

sns.set_theme(color_codes=True)

從dataframe中，獲取y_true(真實標簽)和y_pred(預測標簽)：

y_true = df["target"]
y_pred = df['prediction']

計算驗證數(shù)據(jù)整體的準確率acc、精準率precision、召回率recall、F1，使用加權(quán)模式average=‘weighted’：

# 準確率acc，精準precision，召回recall，F(xiàn)1
acc = accuracy_score(df["target"], df['prediction'])
precision = precision_score(y_true, y_pred, average='weighted')
recall = recall_score(y_true, y_pred, average='weighted')
f1 = f1_score(y_true, y_pred, average='weighted')
print(f'[Info] acc: {acc}, precision: {precision}, recall: {recall}, f1: {f1}')

計算混淆矩陣：

# 橫坐標是真實類別數(shù)，縱坐標是預測類別數(shù)
cf_matrix = confusion_matrix(y_true, y_pred)

5類矩陣的繪制方案，混淆矩陣、百分比的混淆矩陣、召回矩陣、精準矩陣、F1矩陣：

混淆矩陣是計數(shù)，百分比的混淆矩陣是占比
召回矩陣是，每行的和是1，每行代表真實類別數(shù)，占比就是召回
精準矩陣是，每列的和是1，每列代表預測列表數(shù)，占比就是精準
F1矩陣是按照 2PR/(P+R)，注意為0的情況，需要補0，使用np.divide(a, b, out=np.zeros_like(a), where=(b != 0))

代碼如下：

# 橫坐標是真實類別數(shù)，縱坐標是預測類別數(shù)
cf_matrix = confusion_matrix(y_true, y_pred)

figure, axes = plt.subplots(2, 2, figsize=(16*1.25, 16))

# 混淆矩陣
ax = sns.heatmap(cf_matrix, annot=True, fmt='g', ax=axes[0][0], cmap='Blues')
ax.title.set_text("Confusion Matrix")
ax.set_xlabel("y_pred")
ax.set_ylabel("y_true")
# plt.savefig(csv_path.replace(".csv", "_cf_matrix.png"))
# plt.show()

# 混淆矩陣 - 百分比
cf_matrix = confusion_matrix(y_true, y_pred)
ax = sns.heatmap(cf_matrix / np.sum(cf_matrix), annot=True, ax=axes[0][1], fmt='.2%', cmap='Blues')
ax.title.set_text("Confusion Matrix (percent)")
ax.set_xlabel("y_pred")
ax.set_ylabel("y_true")
# plt.savefig(csv_path.replace(".csv", "_cf_matrix_p.png"))
# plt.show()

# 召回矩陣，行和為1
sum_true = np.expand_dims(np.sum(cf_matrix, axis=1), axis=1)
precision_matrix = cf_matrix / sum_true
ax = sns.heatmap(precision_matrix, annot=True, fmt='.2%', ax=axes[1][0], cmap='Blues')
ax.title.set_text("Precision Matrix")
ax.set_xlabel("y_pred")
ax.set_ylabel("y_true")
# plt.savefig(csv_path.replace(".csv", "_recall.png"))
# plt.show()

# 精準矩陣，列和為1
sum_pred = np.expand_dims(np.sum(cf_matrix, axis=0), axis=0)
recall_matrix = cf_matrix / sum_pred
ax = sns.heatmap(recall_matrix, annot=True, fmt='.2%', ax=axes[1][1], cmap='Blues')
ax.title.set_text("Recall Matrix")
ax.set_xlabel("y_pred")
ax.set_ylabel("y_true")
# plt.savefig(csv_path.replace(".csv", "_precision.png"))
# plt.show()

# 繪制4張圖
plt.autoscale(enable=False)
plt.savefig(csv_path.replace(".csv", "_all.png"), bbox_inches='tight', pad_inches=0.2)
plt.show()

# F1矩陣
a = 2 * precision_matrix * recall_matrix
b = precision_matrix + recall_matrix
f1_matrix = np.divide(a, b, out=np.zeros_like(a), where=(b != 0))
ax = sns.heatmap(f1_matrix, annot=True, fmt='.2%', cmap='Blues')
ax.title.set_text("F1 Matrix")
ax.set_xlabel("y_pred")
ax.set_ylabel("y_true")
plt.savefig(csv_path.replace(".csv", "_f1.png"))
plt.show()

輸出混淆矩陣、混淆矩陣(百分比)、召回矩陣、精準矩陣：