本笔记主要记录metrics相关的内容，详细内容请参考代码注释，代码本身只使用了Accuracy和Mean。本节的代码基于上篇笔记FashionMnist的代码经过简单修改而来，上篇笔记链接如下：

Tensorflow2.0笔记 - FashionMnist数据集训练-CSDN博客文章浏览阅读339次。本笔记使用FashionMnist数据集，搭建一个5层的神经网络进行训练，并统计测试集的精度。本笔记中FashionMnist数据集是直接下载到本地加载的方式，不涉及用梯子。关于FashionMnist的介绍，请自行百度。https://blog.csdn.net/vivo01/article/details/136921592?spm=1001.2014.3001.5502

#Fashion Mnist数据集本地下载和加载（不用梯子）
#https://blog.csdn.net/scar2016/article/details/115361245 （百度网盘）
#https://blog.csdn.net/weixin_43272781/article/details/110006990 （github）
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics

tf.__version__


#加载fashion mnist数据集
def load_mnist(path, kind='train'):
    import os
    import gzip
    import numpy as np

    """Load MNIST data from `path`"""
    labels_path = os.path.join(path,
                               '%s-labels-idx1-ubyte.gz'
                               % kind)
    images_path = os.path.join(path,
                               '%s-images-idx3-ubyte.gz'
                               % kind)
    with gzip.open(labels_path, 'rb') as lbpath:
        labels = np.frombuffer(lbpath.read(), dtype=np.uint8,
                               offset=8)
    with gzip.open(images_path, 'rb') as imgpath:
        
        images = np.frombuffer(imgpath.read(), dtype=np.uint8,
                               offset=16).reshape(len(labels), 784)
    return images, labels

#预处理数据
def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32)
    x = tf.convert_to_tensor(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)
    y = tf.convert_to_tensor(y, dtype=tf.int32)
    return x, y
#训练数据
train_data, train_labels = load_mnist("./datasets")
print(train_data.shape, train_labels.shape)
#测试数据
test_data, test_labels = load_mnist("./datasets", "t10k")
print(test_data.shape, test_labels.shape)

batch_size = 128

train_db = tf.data.Dataset.from_tensor_slices((train_data, train_labels))
train_db = train_db.map(preprocess).shuffle(10000).batch(batch_size)

test_db = tf.data.Dataset.from_tensor_slices((test_data, test_labels))
test_db = test_db.map(preprocess).batch(batch_size)

train_db_iter = iter(train_db)
sample = next(train_db_iter)
print('Batch:', sample[0].shape, sample[1].shape)

#定义网络模型
model = Sequential([
    #Layer 1: [b, 784] => [b, 256]
    layers.Dense(256, activation=tf.nn.relu),
    #Layer 2: [b, 256] => [b, 128]
    layers.Dense(128, activation=tf.nn.relu),
    #Layer 3: [b, 128] => [b, 64]
    layers.Dense(64, activation=tf.nn.relu),
    #Layer 4: [b, 64] => [b, 32]
    layers.Dense(32, activation=tf.nn.relu),
    #Layer 5: [b, 32] => [b, 10], 输出类别结果
    layers.Dense(10)
])

#编译网络
model.build(input_shape=[None, 28*28])
model.summary()

#进行训练
total_epoches = 5
learn_rate = 0.01

#Metrics统计
#参考资料：https://zhuanlan.zhihu.com/p/42438077
#1. 新建meter
#acc_meter = metrics.Accuracy()
#loss_meter = metrics.Mean()
#2. 更新状态, update_state()
#loss_meter.update_state(loss)
#acc_meter.update_state(y, pred)
#3.获取结果, result()
#print(step, 'loss:', loss_meter.result().numpy())
#print(step, 'Evaluate Acc:', total_correct/total, acc_meter.result().numpy())
#4.清除度量信息，reset_states()
#loss_meter.reset_states()
#acc_meter.reset_states()


#新建准确度和loss度量对象
acc_meter = metrics.Accuracy()
loss_meter = metrics.Mean()

optimizer = optimizers.Adam(learning_rate = learn_rate)
for epoch in range(total_epoches):
    for step, (x,y) in enumerate(train_db):
        with tf.GradientTape() as tape:
            logits = model(x)
            y_onehot = tf.one_hot(y, depth=10)
            #使用交叉熵作为loss
            loss_ce = tf.reduce_mean(tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True))
            #调用update_state更新loss度量信息
            loss_meter.update_state(loss_ce)
        #计算梯度
        grads = tape.gradient(loss_ce, model.trainable_variables)
        #更新梯度
        optimizer.apply_gradients(zip(grads, model.trainable_variables))

        if step % 100 == 0:
            print("Epoch[", epoch, "]: step-", step, "\tloss: ", loss_meter.result().numpy())
            loss_meter.reset_states()

    #使用测试集进行验证
    total_correct = 0
    total_num = 0
    #清除准确度的统计信息
    acc_meter.reset_states()
    for x,y in test_db:
        logits = model(x)
        #使用softmax得到各个类别的概率
        prob = tf.nn.softmax(logits, axis=1)
        #求出概率最大的结果参数位置，作为预测的分类结果
        pred = tf.cast(tf.argmax(prob, axis=1), dtype=tf.int32)
        #比较结果
        correct = tf.equal(pred, y)
        correct = tf.reduce_sum(tf.cast(correct, dtype=tf.int32))
        #计算精度
        total_correct += int(correct)
        total_num += x.shape[0]
        #使用metircs的update_state进行更新
        acc_meter.update_state(y, pred)
    
    acc = total_correct / total_num
    print("Epoch[", epoch, "] Manual Accuracy:", acc, " Metrics Accuracy:", acc_meter.result().numpy())

运行结果：