tensorflow2.0——數據預處理

import tensorflow as tf
from tensorflow.keras import optimizers,layers
# 定義數據預處理函數
def preprocess(x,y):
    x = tf.cast(x,dtype=tf.float32) / 255                   #   將特征數據轉化為float32類型，並縮放到0到1之間
    y = tf.cast(y,dtype=tf.int32)                           #   將標記數據轉化為int32類型
    y = tf.one_hot(y,depth= 10)                             #   將標記數據轉為one_hot編碼
    return x,y

def get_data():
    # 加載手寫數字數據
    mnist = tf.keras.datasets.mnist
    (train_x, train_y), (test_x, test_y) = mnist.load_data()
    #   開始預處理數據
        #   訓練數據
    db = tf.data.Dataset.from_tensor_slices((train_x,train_y))          #   將數據特征與標記組合
    db = db.map(preprocess)                                             #   根據預處理函數對組合數據進行處理
    db = db.shuffle(60000).batch(100)                                   #   將數據按10000行為單位打亂，並以100行為一個整體進行隨機梯度下降
        #   測試數據
    db_test = tf.data.Dataset.from_tensor_slices((test_x,test_y))
    db_test = db_test.map(preprocess)
    db_test = db_test.shuffle(10000).batch(100)
    return db,db_test

#   測試代碼
db,db_test = get_data()             #   獲取訓練和測試數據
#   設置超參
iter = 100
learn_rate = 0.01
#   定義模型和優化器
model = tf.keras.Sequential([
    layers.Dense(512, activation='relu'),
    layers.Dense(256, activation='relu'),           #   全連接
    layers.Dense(10)
])
optimizer = optimizers.SGD(learning_rate=learn_rate)            #   優化器

#   迭代代碼
for i in range(iter):
    for step,(x,y) in enumerate(db):                            #   對每個batch樣本做梯度計算
        # print('x.shape:{},y.shape:{}'.format(x.shape,y.shape))
        with tf.GradientTape() as tape:
            x = tf.reshape(x,(-1,28*28))               #   將28*28展開為784
            out = model(x)
            loss = tf.reduce_mean(tf.square(out-y))
        grads = tape.gradient(loss,model.trainable_variables)               #   求梯度
        grads,_ = tf.clip_by_global_norm(grads,15)                          #   梯度參數進行限幅，防止偏導的nan和無窮大
        optimizer.apply_gradients(zip(grads,model.trainable_variables))     #   優化器進行參數優化
        if step % 100 == 0:
            print('i:{} ,step:{} ,loss:{}'.format(i, step,loss.numpy()))
            #   求准確率
            acc = tf.equal(tf.argmax(out,axis=1),tf.argmax(y,axis=1))
            acc = tf.cast(acc,tf.int8)
            acc = tf.reduce_mean(tf.cast(acc,tf.float32))
            print('acc:',acc.numpy())