03. Tensorflow2 手写数字识别尝鲜

1. 手写数字识别简述

1.1. MNIST

• 7000 images per category
• train/test splitting: 60k vs 10k
  

1.2. Image

• [28,28,1]
• →[784]
  

1.3. Input and Output

1.4. Regression VS Classification

• 𝑦 = 𝑤 ∗ 𝑥 + 𝑏
  • 𝑦 ∈ 𝑅𝑑
• 𝑜𝑢𝑡 = 𝑋@𝑊 + b
  • out: [0.1, 0.8, 0.02, 0.08]
• 𝑝𝑟𝑒𝑑 = 𝑎𝑟𝑔𝑚𝑎𝑥(𝑜𝑢𝑡)
  • pred: 1
  • pred: 2

1.5. Particularly

1.5. Compute Loss

• 𝑜𝑢𝑡: [1, 10]
• Y/label: 0~9
  • eg.: 1 → [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
  • eg.: 3 → [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0]
• Euclidean Distance（欧式距离）: 𝑜𝑢𝑡 → Label
  • MSE

1.6. In a nutshell

• 𝑜𝑢𝑡 = 𝑟𝑒𝑙𝑢{𝑟𝑒𝑙𝑢 𝑟𝑒𝑙𝑢 𝑋@𝑊1 + 𝑏1 @𝑊2 + 𝑏2 @𝑊3 + 𝑏3}
• 𝑝𝑟𝑒𝑑 = 𝑎𝑟𝑔𝑚𝑎𝑥(𝑜𝑢𝑡)
• 𝑙𝑜𝑠𝑠 = 𝑀𝑆𝐸(𝑜𝑢𝑡, 𝑙𝑎𝑏𝑒𝑙)
• minimize 𝑙𝑜𝑠𝑠
  • [𝑊1, 𝑏1′, 𝑊2′, 𝑏2′, 𝑊3′ , 𝑏3′ ]

2. We need TensorFlow

• NVIDIA TENSOR CORES

3. Code

• X and Y
• 𝑜𝑢𝑡 = 𝑟𝑒𝑙𝑢{𝑟𝑒𝑙𝑢 𝑟𝑒𝑙𝑢 𝑋@𝑊1 + 𝑏1 @𝑊2 + 𝑏2 @𝑊3 + 𝑏3}
• Compute out&loss
• Compute gradient and optimize
• Loop

import os
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers, optimizers, datasets

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

# X and Y
(x, y), (X_val, y_val) = datasets.mnist.load_data()
x = tf.convert_to_tensor(x, dtype=tf.float32) / 255
y = tf.convert_to_tensor(y, dtype=tf.int32)
y = tf.one_hot(y, depth=10)
print(x.shape, y.shape)
train_dataset = tf.data.Dataset.from_tensor_slices((x, y))
train_dataset = train_dataset.batch(200)

# for step, (x, y) in enumerate(train_dataset):
#     print(step, x.shape, y, y.shape)

# 𝑜𝑢𝑡 = 𝑟𝑒𝑙𝑢{𝑟𝑒𝑙𝑢 𝑟𝑒𝑙𝑢 𝑋@𝑊1 + 𝑏1 @𝑊2 + 𝑏2 @𝑊3 + 𝑏3}
model = keras.Sequential([
    layers.Dense(512, activation='relu'),
    layers.Dense(256, activation='relu'),
    layers.Dense(10)
])

optimizer = optimizers.SGD(learning_rate=0.001)


def train_epoch(epoch):
    # loop
    for step, (x, y) in enumerate(train_dataset):
        with tf.GradientTape() as tape:
            # [b, 28, 28] => [b, 784]
            x = tf.reshape(x, (-1, 28 * 28))
            # compute output
            # [b, 784] => [b, 10]
            out = model(x)
            # compute loss
            loss = tf.reduce_sum(tf.square(out - y)) / x.shape[0]

        # optimize and update w1, w2, w3, b1, b2, b3
        grads = tape.gradient(loss, model.trainable_variables)
        # w' = w - lr * grad
        optimizer.apply_gradients((zip(grads, model.trainable_variables)))

        if step % 100 == 0:
            print(epoch, step, 'loss:', loss.numpy())


def train():
    for epoch in range(300):
        train_epoch(epoch)


if __name__ == '__main__':
    train()

4. Pycharm无法对TensorFlow2.0中的Keras模块进行自动补全的解决办法

• TensorFlow2.0之后，Keras作为其主要模块被官方推荐使用。但是在pycharm中使用tensorflow.keras总是无法弹出自动补全，导致编码效率过慢。其实原因很简单，这就是Pycharm的一个bug。下载使用Pycharm2019.3之后的版本就能解决。
• 当前（2019年11月21日）2019.3版本还没有正式发布，可以使用Pycharm EAP版本，也就是测试版，链接在此：点我下载

5. 需要全套课程视频+PPT+代码资源可以私聊我

• 方式1：CSDN私信我!
• 方式2：QQ邮箱:594042358@qq.com或者直接加我QQ: 594042358!