fashion_mnist 计算准确率、召回率、F1值
1、定义
首先需要明确几个概念:
假设某次预测结果统计为下图:
那么各个指标的计算方法为:
- A类的准确率:TP1/(TP1+FP5+FP9+FP13+FP17) 即预测为A的结果中,真正为A的比例
- A类的召回率:TP1/(TP1+FP1+FP2+FP3+FP4) 即实际上所有为A的样例中,能预测出来多少个A(的比例)
- A类的F1值:(准确率*召回率*2)/(准确率+召回率)
实际上我们在训练出某个模型后,会将测试集中每个测试样例进行一次结果预测,因此只需统计这些结果,经过计算即可得到各类数据的准确率、召回率、F1值
2、使用fashion_mnist
需要提前pip安装tensorflow、prettytable、numpy
from tensorflow import keras
import numpy as np
import prettytable
# 下载数据集
fashion_mnist = keras.datasets.fashion_mnist
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()
# 制作标签名称
class_names = ['T-shirt', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Boot']
# 图片数据归一化
train_images = train_images / 255.0
test_images = test_images / 255.0
# 构建3层DNN模型,使用激活函数softmax
model = keras.Sequential([
keras.layers.Flatten(input_shape=(28, 28)),
keras.layers.Dense(128, activation='relu'),
keras.layers.Dense(10, activation='softmax')
])
# 定义模型的损失函数,优化器与评估指标
model.compile(
optimizer=keras.optimizers.Adam(learning_rate=0.001),
loss=keras.losses.sparse_categorical_crossentropy,
metrics=['accuracy']
)
# 训练模型
model.fit(train_images, train_labels, epochs=5)
# 评估模型
predictions = model.predict(test_images)
train_result = np.zeros((10, 10), dtype=int)
for i in range(10000):
train_result[test_labels[i]][np.argmax(predictions[i])] += 1
result_table = prettytable.PrettyTable()
result_table.field_names = ['Type', 'Accu', 'Recall', 'F1']
for i in range(10):
ac = train_result[i][i] / sum(train_result.T[i])
rc = train_result[i][i] / sum(train_result[i])
result_table.add_row([class_names[i], round(ac, 3), round(rc, 3), round(ac * rc * 2 / (ac + rc), 3)])
print(result_table)
实际效果: