前向传播（张量）TensorFlow2.0

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

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

""" 加载数据集
    x: (60000, 28, 28)
    y: (60000,)
"""
(x, y), _ = datasets.mnist.load_data()

# 
x = tf.convert_to_tensor(x, dtype=tf.float32) / 255. # 数据在0~1更利于tf优化
y = tf.convert_to_tensor(y, dtype=tf.int32)

# x、y的范围：
# tf.reduce_min(x), tf.reduce_max(x)
# Out[7]: 
# (<tf.Tensor: id=33, shape=(), dtype=float32, numpy=0.0>,
#  <tf.Tensor: id=35, shape=(), dtype=float32, numpy=1.0>)
# tf.reduce_min(y), tf.reduce_max(y)
# Out[8]: 
# (<tf.Tensor: id=20, shape=(), dtype=uint8, numpy=0>,
# <tf.Tensor: id=27, shape=(), dtype=uint8, numpy=9>)

""" 创建数据集 """
# 这样就可以一次取一个batch
train_db = tf.data.Dataset.from_tensor_slices((x,y)).batch(128)
train_iter = iter(train_db)  # 迭代器
sample = next(train_iter)

# sample[0].shape, sample[1].shape
# Out[13]: (TensorShape([128, 28, 28]), TensorShape([128]))

# [b, 784] => [b, 256] => [b, 128] => [b, 10]
# w: [dim_in, dim_out], b: [dim_out]
''' 若不用tf.Variable包装，则为tf.tensor类型，tf.GradientTape不会跟踪 
    在这个问题里不设置stddev会出现loss的值为nan的情况，即梯度爆炸'''

w1 = tf.Variable(tf.random.truncated_normal([784, 256], stddev=0.1))
b1 = tf.Variable(tf.zeros([256]))
w2 = tf.Variable(tf.random.truncated_normal([256, 128], stddev=0.1))
b2 = tf.Variable(tf.zeros([128]))
w3 = tf.Variable(tf.random.truncated_normal([128, 10], stddev=0.1))
b3 = tf.Variable(tf.zeros([10]))

lr = 1e-3

for epoch in range(10):
    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])
        '''
        tf.GradientTape默认只会跟踪tf.Variable类型的梯度信息,用tf.tensor
        计算出来的grad是None'''
        with tf.GradientTape() as tape: 
            # x: [b, 28*28]
            # h1 = x@w1 + b1
            # [b, 784]@[784, 256] + [256] => [b, 256] + [256] => [b, 256] + [b, 256]
            # TensorFlow会自动完成张量扩张tf.broadcast_to()
            h1 = x@w1 + tf.broadcast_to(b1, [x.shape[0], 256])
            h1 = tf.nn.relu(h1)
            # [b, 256] => [b, 128]
            h2 = h1@w2 + b2
            h2 = tf.nn.relu(h2)
            # [b, 128] => [b, 10]
            out = h2@w3 + b3
    
            # compute loss
            # out: [b, 10]
            # y: [b] => [b, 10]
            y_onehot = tf.one_hot(y, depth=10)
    
            # mse = mean(sum(y-out)^2)
            # [b, 10]
            loss = tf.square(y_onehot - out)
            # mean: scalar
            loss = tf.reduce_mean(loss)
            
        # compute gradients
        grads = tape.gradient(loss, [w1, b1, w2, b2, w3, b3])
        # print(grads)
        # optimizer的作用：
        # w1 = w1 - lr * w1_grad
        '''
        grad为tf.Variable类型，使用 w1 = w1 - lr * grads[0]会让w1从tf.Variable
        变成tf.tensor类型，在再次运算的时候会报错，要使用assign_sub()原地更新，
        w1的数据类型才不会改变'''
    
        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])
        
        # print(isinstance(b3, tf.Variable))
        # print(isinstance(b3, tf.Tensor))
        
        if step % 100 == 0:
            print(epoch, step, 'loss:', float(loss))

结果：

0 0 loss: 0.6338542699813843
0 100 loss: 0.19527742266654968
0 200 loss: 0.19111016392707825
0 300 loss: 0.16401150822639465
0 400 loss: 0.17101189494132996
1 0 loss: 0.1488364338874817
1 100 loss: 0.13843022286891937
1 200 loss: 0.1527971774339676
1 300 loss: 0.13737797737121582
1 400 loss: 0.14406141638755798
2 0 loss: 0.1268216073513031
2 100 loss: 0.12029824405908585
2 200 loss: 0.13270306587219238
2 300 loss: 0.12179452180862427
2 400 loss: 0.12749424576759338
3 0 loss: 0.11288430541753769
3 100 loss: 0.10880843549966812
3 200 loss: 0.11929886043071747
3 300 loss: 0.11120941489934921
3 400 loss: 0.11638245731592178
4 0 loss: 0.10308767855167389
4 100 loss: 0.10074740648269653
4 200 loss: 0.10969498008489609
4 300 loss: 0.10343489795923233
4 400 loss: 0.10824362933635712
5 0 loss: 0.09583660215139389
5 100 loss: 0.09478463232517242
5 200 loss: 0.10255211591720581
5 300 loss: 0.09745050221681595
5 400 loss: 0.1019618883728981
...

若不设置sttdev，则容易梯度爆炸：

0 0 loss: 387908.8125
0 100 loss: nan
0 200 loss: nan
0 300 loss: nan
0 400 loss: nan
1 0 loss: nan
1 100 loss: nan
1 200 loss: nan
1 300 loss: nan
1 400 loss: nan
2 0 loss: nan
2 100 loss: nan
2 200 loss: nan
2 300 loss: nan
2 400 loss: nan
3 0 loss: nan
3 100 loss: nan
3 200 loss: nan
3 300 loss: nan
3 400 loss: nan
...