TensorFlow2模型训练框架MNIST为例

step1、数据集预处理

import  tensorflow as tf

from tensorflow import keras
from tensorflow.keras import layers, Sequential,datasets,optimizers

# 将数据转换为tensor
def preprocess(x, y):
    # [0~255] => [-1~1]
    x = 2 * tf.cast(x, dtype=tf.float32) / 255.  -  1.
    y = tf.cast(y, dtype=tf.int32)
    return x,y
# 数据加载
(x, y), (x_val, y_val) = datasets.mnist.load_data()
# 制作训练集以及测试集，方便迭代获取每个batch
batchsz = 128
train_db = tf.data.Dataset.from_tensor_slices((x,y))
train_db = train_db.map(preprocess).shuffle(10000).batch(batchsz)
test_db = tf.data.Dataset.from_tensor_slices((x_val, y_val))
test_db = test_db.map(preprocess).batch(batchsz)

step2、模型构建

model = Sequential([ 
    layers.Dense(512, activation='relu'),
    layers.Dense(256, activation='relu'),
    layers.Dense(10)])
model.build(input_shape=[None, 28*28])
optimizer = optimizers.SGD(learning_rate=1e-3)

step3、模型训练

for epoch in range(5):
    
    for step,(x,y) in enumerate(train_db ):
        # x: [b, 28, 28] => [b, 784]
        # y: [b]
        x = tf.reshape(x, [-1, 28*28])
        
        with tf.GradientTape() as tape:
            # [b, 784] => [b, 10]
            logits = model(x) # 最后一层的输出，该层未设置激活函数
            y_onehot = tf.one_hot(y, depth=10)
            # 计算损失，得到标量[b]
            loss_mse = tf.reduce_mean(tf.losses.MSE(y_onehot, logits)) # 以MSE做损失函数
            # 以交叉熵做损失函数，用于分类
            loss_ce = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
            loss_ce = tf.reduce_mean(loss_ce)
        
        # 利用交叉熵损失函数来求梯度
        # model.trainable_variables包含所有需更新参数
        grads = tape.gradient(loss_ce, model.trainable_variables)
        # 利用梯度对参数进行更新
        optimizer.apply_gradients(zip(grads, model.trainable_variables)) 
        
        if step % 100 == 0:
                print(epoch, step, 'loss:', float(loss_ce), float(loss_mse))
    
    # 在测试集上进行测试
    total_correct = 0
    total_num = 0
    for x, y in test_db :
        # x: [b, 28, 28] => [b, 784]
        # y: [b]
        x = tf.reshape(x, [-1, 28*28])
        # [b, 10]
        logits = model(x)
        # logits => prob, [b, 10]
        prob = tf.nn.softmax(logits, axis=1)
        # [b, 10] => [b], int64
        pred = tf.argmax(prob, axis=1) # 得到概率值最大的索引
        pred = tf.cast(pred, dtype=tf.int32) 
        # pred:[b]
        # y: [b]
        # correct: [b], True: equal, False: not equal
        # 统计 True 的个数，即预测正确的个数
        correct = tf.equal(pred, y)
        correct = tf.reduce_sum(tf.cast(correct, dtype=tf.int32))

        total_correct += int(correct)
        total_num += x.shape[0]

    acc = total_correct / total_num
    print(epoch, 'test acc:', acc)