TensorFlow2-前向传播(张量)-实战(10)

前向传播(张量)-实战

What we have learned

• create tensor 创建张量
• indexing and slices 索引和切片
• reshape and broadcasting 重塑与传播
• math operations 数学运算

Recap

• 𝑜𝑢𝑡 = 𝑟𝑒𝑙𝑢{𝑟𝑒𝑙𝑢{𝑟𝑒𝑙𝑢{[𝑋@𝑊1 + 𝑏1]}@𝑊2 + 𝑏2}@𝑊3 + 𝑏3}
• 𝑝𝑟𝑒𝑑 = 𝑎𝑟𝑔𝑚𝑎𝑥(𝑜𝑢𝑡)
• 𝑙𝑜𝑠𝑠 = 𝑀𝑆𝐸(𝑜𝑢𝑡, 𝑙𝑎𝑏𝑒𝑙)
• minimize 𝑙𝑜𝑠𝑠
  • $[W_1^{{}^{\prime }},b_1^{{}^{\prime }},W_2^{{}^{\prime }},b_2^{{}^{\prime }},W_3^{{}^{\prime }},b_3^{{}^{\prime }}]$

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

# 日志级别
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# 加载数据集
# x: [60k, 28, 28],
# y: [60k]
(x, y), _ = datasets.mnist.load_data()

# 数据类型转换为Tensor
x = tf.convert_to_tensor(x, dtype=tf.float32)
y = tf.convert_to_tensor(y, dtype=tf.int32)

# 查看数据形状，数据类型
print(x.shape, y.shape, x.dtype, y.dtype)

(60000, 28, 28) (60000,) <dtype: 'float32'> <dtype: 'int32'>

# 查看最大值、最小值
print(tf.reduce_min(x), tf.reduce_max(x)) # 0-255
print(tf.reduce_min(y), tf.reduce_max(y)) # 0-9

tf.Tensor(0.0, shape=(), dtype=float32) tf.Tensor(255.0, shape=(), dtype=float32)
tf.Tensor(0, shape=(), dtype=int32) tf.Tensor(9, shape=(), dtype=int32)

# x：数据缩放
# x: [0~255] => [0~1.]
x = tf.convert_to_tensor(x, dtype=tf.float32) / 255.0

# 查看最大值、最小值
print(tf.reduce_min(x), tf.reduce_max(x)) # 0-1.

tf.Tensor(0.0, shape=(), dtype=float32) tf.Tensor(1.0, shape=(), dtype=float32)

# 将 tensor 转换为 datasets
# tensorflow 推荐使用 tf.data.Dataset 来加载数据集
# tensor => datasets
# 128个为一个batch返回datasets
train_db = tf.data.Dataset.from_tensor_slices((x, y)).batch(128)

# 测试迭代数据
train_iter = iter(train_db)
sample = next(train_iter)
print('batch:', sample[0].shape, sample[1].shape)
sample = next(train_iter)
print('batch:', sample[0].shape, sample[1].shape)

batch: (128, 28, 28) (128,)
batch: (128, 28, 28) (128,)

# [b, 784] => [b, 256] => [b, 128] => [b, 10]
# [dim_in, dim_out], [dim_out]
# 
w1 = tf.random.truncated_normal([784, 256], stddev=0.1)
b1 = tf.zeros([256])
w2 = tf.random.truncated_normal([256, 128], stddev=0.1)
b2 = tf.zeros([128])
w3 = tf.random.truncated_normal([128, 10], stddev=0.1)
b3 = tf.zeros([10])
b3 # tf.Tensor

<tf.Tensor: shape=(10,), dtype=float32, numpy=array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)>

tf.GradientTape()

• 记录自动区分操作。
• 默认跟踪类型：tf.Variable
• 所以数据需要包装

# [b, 784] => [b, 256] => [b, 128] => [b, 10]
# [dim_in, dim_out], [dim_out]
# stddev=0.1  解决梯度爆炸的问题
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):  # iterate db for 10
    for step, (x, y) in enumerate(train_db):  # for every batch
        # x:[128, 28, 28]
        # y: [128]

        # [b, 28, 28] => [b, 28*28]
        # b: x.shape[0]
        x = tf.reshape(x, [-1, 28 * 28])
        
        # 记录自动区分操作。
        # 默认跟踪类型：tf.Variable
        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]
            h1 = x @ w1 + tf.broadcast_to(b1, [x.shape[0], 256])
            h1 = tf.nn.relu(h1)
            
            # [b, 256] => [b, 128]
            # b2: 会自动转broadcast_to
            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 计算坡度
        # 使用此磁带上下文中记录的操作计算渐变。
        grads = tape.gradient(loss, [w1, b1, w2, b2, w3, b3])
        # print(grads)
        
        # 重新赋值，数据类型变化
        # w1 = w1 - lr * w1_grad
        
        # assign_sub 实现原地更新操作
        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))

0 0 loss: 0.4980630874633789
0 100 loss: 0.20007649064064026
0 200 loss: 0.1879170686006546
0 300 loss: 0.1585090458393097
0 400 loss: 0.17265315353870392
1 0 loss: 0.15226225554943085
1 100 loss: 0.14862820506095886
1 200 loss: 0.15506379306316376
1 300 loss: 0.13298840820789337
1 400 loss: 0.14317986369132996
2 0 loss: 0.12775573134422302
2 100 loss: 0.13045187294483185
2 200 loss: 0.13609960675239563
2 300 loss: 0.11735637485980988
2 400 loss: 0.1252080500125885
3 0 loss: 0.11262925714254379
3 100 loss: 0.11829018592834473
3 200 loss: 0.12328527122735977
3 300 loss: 0.10690398514270782
3 400 loss: 0.11292298138141632
4 0 loss: 0.10224197059869766
4 100 loss: 0.10958822816610336
4 200 loss: 0.11386322975158691
4 300 loss: 0.0994718074798584
4 400 loss: 0.10404429584741592
5 0 loss: 0.09465127438306808
5 100 loss: 0.10307874530553818
5 200 loss: 0.10661985725164413
5 300 loss: 0.09376838058233261
5 400 loss: 0.09735797345638275
6 0 loss: 0.08888636529445648
6 100 loss: 0.09798622876405716
6 200 loss: 0.1008676290512085
6 300 loss: 0.08918877691030502
6 400 loss: 0.09213654696941376
7 0 loss: 0.08434510231018066
7 100 loss: 0.09376645088195801
7 200 loss: 0.0961538702249527
7 300 loss: 0.08544725924730301
7 400 loss: 0.08790375292301178
8 0 loss: 0.0806211456656456
8 100 loss: 0.09019575268030167
8 200 loss: 0.09219495952129364
8 300 loss: 0.0822930559515953
8 400 loss: 0.08442799001932144
9 0 loss: 0.07753437757492065
9 100 loss: 0.08713443577289581
9 200 loss: 0.08879637718200684
9 300 loss: 0.07959921658039093
9 400 loss: 0.08146005868911743