神经网络与深度学习（邱锡鹏）编程练习4 FNN 简单神经网络 Jupyter导出版 TensorFlow

GitHub - nndl/nndl-exercise-ans: Solutions for nndl/exercise

准备数据

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

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'  # or any {'0', '1', '2'}

def mnist_dataset():
    (x, y), (x_test, y_test) = datasets.mnist.load_data()
    #normalize
    x = x/255.0
    x_test = x_test/255.0
    
    return (x, y), (x_test, y_test)

print(list(zip([1, 2, 3, 4], ['a', 'b', 'c', 'd'])))

[(1, 'a'), (2, 'b'), (3, 'c'), (4, 'd')]

建立模型

class myModel:
    def __init__(self):
        ####################
        '''声明模型对应的参数'''
        ####################
        self.W1 = tf.Variable(shape=[28*28, 100], dtype=tf.float32,
                         initial_value=tf.random.uniform(shape=[28*28, 100],
                                                         minval=-0.1, maxval=0.1))
        self.b1 = tf.Variable(shape=[100], dtype=tf.float32, initial_value=tf.zeros(100))
        self.W2 = tf.Variable(shape=[100, 10], dtype=tf.float32,
                         initial_value=tf.random.uniform(shape=[100, 10],
                                                         minval=-0.1, maxval=0.1))
        self.b2 = tf.Variable(shape=[10], dtype=tf.float32, initial_value=tf.zeros(10))
        self.trainable_variables = [self.W1, self.W2, self.b1, self.b2]
    def __call__(self, x):
        ####################
        '''实现模型函数体，返回未归一化的logits'''
        ####################
        flat_x = tf.reshape(x, shape=[-1, 28*28])
        h1 = tf.tanh(tf.matmul(flat_x, self.W1) + self.b1)
        logits = tf.matmul(h1, self.W2) + self.b2
        return logits
        
model = myModel()

optimizer = optimizers.Adam()

计算 loss

@tf.function
def compute_loss(logits, labels):
    return tf.reduce_mean(
        tf.nn.sparse_softmax_cross_entropy_with_logits(
            logits=logits, labels=labels))

@tf.function
def compute_accuracy(logits, labels):
    predictions = tf.argmax(logits, axis=1)
    return tf.reduce_mean(tf.cast(tf.equal(predictions, labels), tf.float32))

@tf.function
def train_one_step(model, optimizer, x, y):
    with tf.GradientTape() as tape:
        logits = model(x)
        loss = compute_loss(logits, y)

    # compute gradient
    trainable_vars = [model.W1, model.W2, model.b1, model.b2]
    grads = tape.gradient(loss, trainable_vars)
    for g, v in zip(grads, trainable_vars):
        v.assign_sub(0.01*g)

    accuracy = compute_accuracy(logits, y)

    # loss and accuracy is scalar tensor
    return loss, accuracy

@tf.function
def test(model, x, y):
    logits = model(x)
    loss = compute_loss(logits, y)
    accuracy = compute_accuracy(logits, y)
    return loss, accuracy

实际训练

train_data, test_data = mnist_dataset()
for epoch in range(50):
    loss, accuracy = train_one_step(model, optimizer, 
                                    tf.constant(train_data[0], dtype=tf.float32), 
                                    tf.constant(train_data[1], dtype=tf.int64))
    print('epoch', epoch, ': loss', loss.numpy(), '; accuracy', accuracy.numpy())
loss, accuracy = test(model, 
                      tf.constant(test_data[0], dtype=tf.float32), 
                      tf.constant(test_data[1], dtype=tf.int64))

print('test loss', loss.numpy(), '; accuracy', accuracy.numpy())

epoch 0 : loss 2.3099253 ; accuracy 0.13093333
epoch 1 : loss 2.304885 ; accuracy 0.13463333
epoch 2 : loss 2.2998776 ; accuracy 0.13906667
epoch 3 : loss 2.2949016 ; accuracy 0.14308333
epoch 4 : loss 2.2899568 ; accuracy 0.14741667
epoch 5 : loss 2.2850416 ; accuracy 0.15208334
epoch 6 : loss 2.2801554 ; accuracy 0.15626666
epoch 7 : loss 2.2752976 ; accuracy 0.1611
epoch 8 : loss 2.2704673 ; accuracy 0.16601667
epoch 9 : loss 2.265663 ; accuracy 0.17066666
epoch 10 : loss 2.2608845 ; accuracy 0.17548333
epoch 11 : loss 2.256132 ; accuracy 0.18088333
epoch 12 : loss 2.2514026 ; accuracy 0.18561667
epoch 13 : loss 2.2466977 ; accuracy 0.19076666
epoch 14 : loss 2.2420156 ; accuracy 0.19616666
epoch 15 : loss 2.2373557 ; accuracy 0.20183334
epoch 16 : loss 2.2327178 ; accuracy 0.2075
epoch 17 : loss 2.228101 ; accuracy 0.21338333
epoch 18 : loss 2.2235048 ; accuracy 0.21935
epoch 19 : loss 2.2189286 ; accuracy 0.22548333
epoch 20 : loss 2.214372 ; accuracy 0.23158333
epoch 21 : loss 2.209834 ; accuracy 0.23836666
epoch 22 : loss 2.205315 ; accuracy 0.24433333
epoch 23 : loss 2.2008135 ; accuracy 0.25088334
epoch 24 : loss 2.1963296 ; accuracy 0.2572
epoch 25 : loss 2.1918633 ; accuracy 0.26376668
epoch 26 : loss 2.187413 ; accuracy 0.27018332
epoch 27 : loss 2.1829789 ; accuracy 0.27626666
epoch 28 : loss 2.1785607 ; accuracy 0.28215
epoch 29 : loss 2.1741576 ; accuracy 0.28843334
epoch 30 : loss 2.1697698 ; accuracy 0.2942
epoch 31 : loss 2.165396 ; accuracy 0.30048335
epoch 32 : loss 2.1610367 ; accuracy 0.30646667
epoch 33 : loss 2.1566913 ; accuracy 0.31136668
epoch 34 : loss 2.152359 ; accuracy 0.31738332
epoch 35 : loss 2.1480403 ; accuracy 0.32326666
epoch 36 : loss 2.1437342 ; accuracy 0.32873333
epoch 37 : loss 2.1394403 ; accuracy 0.33388335
epoch 38 : loss 2.1351585 ; accuracy 0.33935
epoch 39 : loss 2.1308892 ; accuracy 0.34543332
epoch 40 : loss 2.1266308 ; accuracy 0.35115
epoch 41 : loss 2.1223838 ; accuracy 0.35598335
epoch 42 : loss 2.1181479 ; accuracy 0.36223334
epoch 43 : loss 2.1139226 ; accuracy 0.36766666
epoch 44 : loss 2.109708 ; accuracy 0.37271667
epoch 45 : loss 2.1055036 ; accuracy 0.37828332
epoch 46 : loss 2.101309 ; accuracy 0.38358334
epoch 47 : loss 2.097124 ; accuracy 0.38878334
epoch 48 : loss 2.092949 ; accuracy 0.39408332
epoch 49 : loss 2.088783 ; accuracy 0.39916667
test loss 2.0848043 ; accuracy 0.4049