深度学习与TensorFlow 2入门实战——前向传播实战，代码详解

import  tensorflow as tf
from    tensorflow import keras
from    tensorflow.keras import datasets
import  os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' # 有助于减少tf环境构建时的废话

# x: [60k, 28, 28],
# y: [60k]
(x, y), _ = datasets.mnist.load_data() # 自动下载数据集
# x: [0~255] => [0~1.]
x = tf.convert_to_tensor(x, dtype=tf.float32) / 255. # 将numpy数据转化为tensor数据
y = tf.convert_to_tensor(y, dtype=tf.int32)

print(x.shape, y.shape, x.dtype, y.dtype) # 检查数据的shape和类型是否与预期一致
print(tf.reduce_min(x), tf.reduce_max(x)) # 输出x和y的最小值和最大值
print(tf.reduce_min(y), tf.reduce_max(y))

# 输出：
# (60000, 28, 28) (60000,) <dtype: 'float32'> <dtype: 'int32'>
# tf.Tensor(0.0, shape=(), dtype=float32) tf.Tensor(1.0, shape=(), dtype=float32)
# tf.Tensor(0, shape=(), dtype=int32) tf.Tensor(9, shape=(), dtype=int32)


train_db = tf.data.Dataset.from_tensor_slices((x,y)).batch(128) # 创建数据集对象
train_iter = iter(train_db) # 创建迭代器
sample = next(train_iter) # 用next对迭代器不停迭代，直到获得最终数据
print('batch:', sample[0].shape, sample[1].shape) # 输出batch中第一和第二个元素
# 输出：batch (128, 28, 28) (128,)

# 降维[b, 784] => [b, 256] => [b, 128] => [b, 10]
# [dim_in, dim_out], [dim_out]
w1 = tf.Variable(tf.random.truncated_normal([784, 256], stddev=0.1)) # 用variable包装权值，使之在之后可以进行求导训练
b1 = tf.Variable(tf.zeros([256]))
w2 = tf.Variable(tf.random.truncated_normal([256, 128], stddev=0.1)) # 截断正态分布
b2 = tf.Variable(tf.zeros([128]))
w3 = tf.Variable(tf.random.truncated_normal([128, 10], stddev=0.1))
b3 = tf.Variable(tf.zeros([10]))
# tf.truncated_normal(shape, mean, stddev, dtype, seed, name)
#shape：表示生成随机数的维度
# mean：正太分布的均值，默认为0
# stddev：正太分布的标准差
# dtype：生成正太分布数据的类型
# seed：一个整数，当设置之后，每次生成的随机数都一样
# name：正太分布的名字
# tf.truncated_normal([2, 2], mean=0, stddev=0.2, dtype=tf.float32, seed=1, name='v')
# 均值mean=0，stddev=0.2，则生成的随机数与均值差不能大于两倍中误差，即范围为：[-0.4,0.4]
# tf.truncated_normal([2, 2], mean=0, stddev=0.1, dtype=tf.float32, seed=1, name='v')
# 设置stddev=0.1，则生成的范围：[-0.2,0.2]

lr = 1e-3 # 表示梯度步长（学习速率）

for epoch in range(10): # iterate db for 10，从0开始，到9结束的10个数，迭代十次
    for step, (x, y) in enumerate(train_db): # for every batch
        # x:[128, 28, 28]
        # y: [128]

        # [b, 28, 28] => [b, 28*28]
        x = tf.reshape(x, [-1, 28*28]) # 这里的x的shape为[b,28*28]，从[b,w,h]变成[b,w*h]
        # 对输入特征项的维度变换，-1会自动计算b
        with tf.GradientTape() as tape: # 自动求导计算梯度，只会跟踪tf.Variable类型数据
            # x: [b, 28*28]
            # h1 = x@w1 + b1
            # x[b, 784]@w1[784, 256] + b1[256] => h1[b, 256] + b2[256] => h2[b, 256] + b3[b, 256]
            h1 = x@w1 + tf.broadcast_to(b1, [x.shape[0], 256]) # [b,784] @ [784,256] + [b,256] = [b,256]
            # 激活函数，引入非线性因子
            h1 = tf.nn.relu(h1)
            # [b, 256] => [b, 128]
            h2 = h1@w2 + b2
            h2 = tf.nn.relu(h2) # ==2== 从隐含层1到隐含层2，[b,256] @ [256,128] + [b,128] = [b,128]
            # [b, 128] => [b, 10]
            out = h2@w3 + b3 # ==3== 从隐含层2到输出层，[b,128] @ [128,10] + [b,10] = [b,10]

            # compute loss，计算误差，mse = mean(sum(y-out)^2)
            # out: [b, 10],使y: [b] => y_onehot: [b, 10]
            # 使分类进行独热编码（实际上是一种升维），depth表示类型（标签）数量
            # out: [b, 10]
            # y: [b] => [b, 10]
            y_onehot = tf.one_hot(y, depth=10)

            # mse = mean(sum(y-out)^2)
            # [b, 10]计算误差，输出值out的shape为[b,10]，onehot编码后真实值y的shape为[b,10]
            # 计算均方差 mse = mean(sum((y-out)^2)
            loss = tf.square(y_onehot - out)
            # mean: scalar
            loss = tf.reduce_mean(loss)

        # compute gradients梯度计算
        grads = tape.gradient(loss, [w1, b1, w2, b2, w3, b3])
        # print(grads)
        # w1 = w1 - lr * w1_grad
        # w1 = w1 - lr * grads[0]中易错点：新的w1为Tenor类型，没有经过Variable，不能自动求导

        # 原地更新方法：assign_sub(ref, value)方法:ref = ref - value
        # tensor. assign_sub(value)：功能tensor = tensor - assign_sub
        w1.assign_sub(lr * grads[0])
        b1.assign_sub(lr * grads[1])
        w2.assign_sub(lr * grads[2])
        b2.assign_sub(lr * grads[3])
        w3.assign_sub(lr * grads[4])
        b3.assign_sub(lr * grads[5])

        # 每100个batch计算之后输出loss数值，用float将loss由tensor转换为numpy数值
        if step % 100 == 0:
            print(epoch, step, 'loss:', float(loss))

# 一个epoch代表全部数据进入网络一次，这个时候，整个网络结构只对这批数据全部走完一次
# batch指的是将一个epoch(所有数据)分割成几份，每份的大小为batch size

 代码解释都来自于自己理解和网上资料整合

代码可直接运行，输出结果

(60000, 28, 28) (60000,) <dtype: 'float32'> <dtype: 'int32'>
tf.Tensor(0.0, shape=(), dtype=float32) tf.Tensor(1.0, shape=(), dtype=float32)
tf.Tensor(0, shape=(), dtype=int32) tf.Tensor(9, shape=(), dtype=int32)
batch: (128, 28, 28) (128,)
0 0 loss: 0.33306556940078735
0 100 loss: 0.20801925659179688
0 200 loss: 0.18559986352920532
0 300 loss: 0.16979117691516876
0 400 loss: 0.1634766310453415
1 0 loss: 0.16542527079582214
1 100 loss: 0.1560421586036682
1 200 loss: 0.15167365968227386
1 300 loss: 0.14091989398002625
1 400 loss: 0.13649582862854004
2 0 loss: 0.1405315101146698
2 100 loss: 0.136332705616951
2 200 loss: 0.13225948810577393
2 300 loss: 0.12394670397043228
2 400 loss: 0.1201285719871521
3 0 loss: 0.12452125549316406
3 100 loss: 0.12340305745601654
3 200 loss: 0.11915738880634308
3 300 loss: 0.11255736649036407
3 400 loss: 0.109268918633461
4 0 loss: 0.11321870237588882
4 100 loss: 0.1141565814614296
4 200 loss: 0.1097048670053482
4 300 loss: 0.10436363518238068
4 400 loss: 0.10156302154064178
5 0 loss: 0.10487793385982513
5 100 loss: 0.10716714709997177
5 200 loss: 0.10255954414606094
5 300 loss: 0.098089799284935
5 400 loss: 0.09579628705978394
6 0 loss: 0.09836559742689133
6 100 loss: 0.10161657631397247
6 200 loss: 0.09695516526699066
6 300 loss: 0.0931091457605362
6 400 loss: 0.09125898778438568
7 0 loss: 0.09311607480049133
7 100 loss: 0.0970434620976448
7 200 loss: 0.0923660472035408
7 300 loss: 0.08903666585683823
7 400 loss: 0.08764063566923141
8 0 loss: 0.08880780637264252
8 100 loss: 0.09320847690105438
8 200 loss: 0.08853678405284882
8 300 loss: 0.0856456384062767
8 400 loss: 0.084683857858181
9 0 loss: 0.08519560843706131
9 100 loss: 0.08989918977022171
9 200 loss: 0.08525765687227249
9 300 loss: 0.08276523649692535
9 400 loss: 0.0821666494011879