感觉这个比前一套，容易理解些~~

关于数据提前下载的问题：

https://www.jianshu.com/p/5116046733fe

如果使用keras的cifar10.load_data()函数，你会发现，代码会自动去下载 cifar-10-python.tar.gz 文件
实际上，通过查看cifar10.py和site-packages/keras/utils/data_utils.py的get_file函数，你会发现，代码将将下载后的文件存放在 ～./keras/datasets目录下，但是！！！！文件名却被改成了 cifar-10-batches-py.tar.gz

惊不惊喜，意不意外？所以如果要避免下载，已经有数据集了，应该:
cp cifar-10-python.tar.gz ～./keras/datasets/cifar-10-batches-py.tar.gz

完美解决问题！

作者：不爱吃饭的小孩怎么办
链接：https://www.jianshu.com/p/5116046733fe
来源：简书
简书著作权归作者所有，任何形式的转载都请联系作者获得授权并注明出处。

import timeit
import tensorflow as tf
import numpy as np
from tensorflow.keras.datasets.cifar10 import load_data

def model():
x = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
y = tf.placeholder(tf.float32, shape=[None, 10])
rate = tf.placeholder(tf.float32)
# convolutional layer 1
conv_1 = tf.layers.conv2d(x, 32, [3, 3], padding='SAME', activation=tf.nn.relu)
max_pool_1 = tf.layers.max_pooling2d(conv_1, [2, 2], strides=2, padding='SAME')
drop_1 = tf.layers.dropout(max_pool_1, rate=rate)
# convolutional layer 2
conv_2 = tf.layers.conv2d(drop_1, 64, [3, 3], padding="SAME", activation=tf.nn.relu)
max_pool_2 = tf.layers.max_pooling2d(conv_2, [2, 2], strides=2, padding="SAME")
drop_2 = tf.layers.dropout(max_pool_2, rate=rate)
# convolutional layers 3
conv_3 = tf.layers.conv2d(drop_2, 128, [3, 3], padding="SAME", activation=tf.nn.relu)
max_pool_3 = tf.layers.max_pooling2d(conv_3, [2, 2], strides=2, padding="SAME")
drop_3 = tf.layers.dropout(max_pool_3, rate=rate)
# fully connected layer 1
flat = tf.reshape(drop_3, shape=[-1, 4 * 4 * 128])
fc_1 = tf.layers.dense(flat, 80, activation=tf.nn.relu)
drop_4 = tf.layers.dropout(fc_1 , rate=rate)
# fully connected layer 2 or the output layers
fc_2 = tf.layers.dense(drop_4, 10)
output = tf.nn.relu(fc_2)
# accuracy
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(output, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# loss
loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=output, labels=y))
# optimizer
optimizer = tf.train.AdamOptimizer(1e-4, beta1=0.9, beta2=0.999, epsilon=1e-8).minimize(loss)
return x, y, rate, accuracy, loss, optimizer

def one_hot_encoder(y):
ret = np.zeros(len(y) * 10)
ret = ret.reshape([-1, 10])
for i in range(len(y)):
ret[i][y[i]] = 1
return (ret)

def train(x_train, y_train, sess, x, y, rate, optimizer, accuracy, loss):
batch_size = 128
y_train_cls = one_hot_encoder(y_train)
start = end = 0
for i in range(int(len(x_train) / batch_size)):
if (i + 1) % 100 == 1:
start = timeit.default_timer()
batch_x = x_train[i * batch_size:(i + 1) * batch_size]
batch_y = y_train_cls[i * batch_size:(i + 1) * batch_size]
_, batch_loss, batch_accuracy = sess.run([optimizer, loss, accuracy], feed_dict={x:batch_x, y:batch_y, rate:0.4})
if (i + 1) % 100 == 0:
end = timeit.default_timer()
print("Time:", end-start, "s the loss is ", batch_loss, " and the accuracy is ", batch_accuracy * 100, "%")

def test(x_test, y_test, sess, x, y, rate, accuracy, loss):
batch_size = 64
y_test_cls = one_hot_encoder(y_test)
global_loss = 0
global_accuracy = 0
for t in range(int(len(x_test) / batch_size)):
batch_x = x_test[t * batch_size : (t + 1) * batch_size]
batch_y = y_test_cls[t * batch_size : (t + 1) * batch_size]
batch_loss, batch_accuracy = sess.run([loss, accuracy], feed_dict={x:batch_x, y:batch_y, rate:1})
global_loss += batch_loss
global_accuracy += batch_accuracy
global_loss = global_loss / (len(x_test) / batch_size)
global_accuracy = global_accuracy / (len(x_test) / batch_size)
print("In Test Time, loss is ", global_loss, ' and the accuracy is ', global_accuracy)

EPOCH = 100
(x_train, y_train), (x_test, y_test) = load_data()
print("There is ", len(x_train), " training images and ", len(x_test), " images")
x, y, rate, accuracy, loss, optimizer = model()
sess = tf.Session()
sess.run(tf.global_variables_initializer())

for i in range(EPOCH):
print("Train on epoch ", i ," start")
train(x_train, y_train, sess, x, y, rate, optimizer, accuracy, loss)
test(x_train, y_train, sess, x, y, rate, accuracy, loss)