Tensorflow 学习笔记

# 在用tensorflow前先打开
$ source ~/tensorflow/bin/activate

# 激活后会有标识
(tensorflow)$

#用完后关闭
(tensorflow)$ deactivate

# 删除tensorflow
$ rm -r ~/tensorflow

1.  Getting Started With Tensorflow

　　Tensorflow的基本单元是张量(tensor)，张量的阶(rank)是它的维度(the number of dimension)。

　　

　　Tensorflow由一系列运算图(computational graph)组成。点有多种类型，包括常数(constant)、变量(variable)。

　　

　　要操作computational graph中的node，要用到session。

sess = tf.Session()
print(sess.run([node1, node2]))

　　可视化： TensorBoard可以用来表现computational graph。

　　

　　用placeholder可以完成类似函数(function)的功能

a = tf.placeholder(tf.float32)
b = tf.placeholder(tf.float32)
adder_node = a + b # + provides a shortcut for tf.add(a, b)

print(sess.run(adder_node, {a: 3, b: 4.5}))

　　利用变量(variables)给节点添加参数

2. Deep MNIST for Experts

InteractiveSession class is more convenient and flexible.

import tensorflow as tf
sess = tf.InteractiveSession()

Placeholders with shape argument can automatically catch bugs regarding on manipulation on inconsistent tensor shape.

x = tf.placeholder(tf.float32, shape=[None, 784])  # None means that the number of sample is variable and the length of each sample is 784 (28 * 28)

 包含Deep convolutional neural networks的使用教程。

3. Read Data

batchsize: 每次训练在训练集中取得样本数

iteration: 使用batchsize个样本训练一次

epoch: 使用训练集中的全部样本训练一次

# 读入的数据以列表存储
filename = []
# tf.train.string_input_producer() 产生一个文件名队列
filename_queue = tf.train.string_input_producer(filename, shuffle=False, num_epochs=5）# reader从文件名队列中读数据。对应的方法是reader.read
reader = tf.WholeFileReader()
key, value = reader.read(filename_queue)
# tf.train.string_input_producer定义了一个epoch变量，要对它进行初始化
tf.local_variables_initializer().run()
# 使用start_queue_runners之后，才会开始填充队列
threads = tf.train.start_queue_runners(sess=sess)

详见: http://blog.csdn.net/buptgshengod/article/details/72956846

numpy

import numpy as np
np.loadtext(c)  # return ndarray
np.ones()  # return a new array of given shape and type, filled with ones
i = 1 * 10 ** (-10)  # ** means power, and its priority is higher than *
np.eye(3, dtype = int)  # diagonal matrix, consist of 0 and 1
vector = np.hstack((vector, temp))  # concatenation along the second axis (horizontally in 2 dimension)
np.save(filename, array, fmt = '%d')  # save the data of int format

4. make hand dirty

numpy一位数组转置只需要操作shape

array_1d = np.array([1, 2])  
array_1d.shape = (2, 1)  # finished

array_2d.transpose()  # 二维数组的转置方法

python assert断言是声明其布尔值必须为真的判定

assert len(lists) >=5  # 若为假则触发异常

 使用名称前的单下划线，用于指定该名称属性为“私有”，其只供内部使用。 （_spam)

 collections是一个集合模块，提供许多有用的集合类，如namedtuple(提供一个列表的集合类）

>>> from collections import namedtuple
>>> Point = namedtuple('Point', ['x', 'y'])
>>> p = Point(1, 2)
>>> p.x
1
>>> p.y
2

 5. Model　　

　　strides 中的四个参数，第一个是 the number of images，第二个是 the height of images，第三个是 the width of images，第四个是 the number of channel


　　小白学Tensorflow之卷积神经网络 https://www.jianshu.com/p/70c8d6663b00

　　深入MNIST http://www.tensorfly.cn/tfdoc/tutorials/mnist_pros.html

　　Batch Normalization 批标准化 https://morvanzhou.github.io/tutorials/machine-learning/tensorflow/5-13-BN/

　　martin-gorner/tensorflow-mnist-tutorial (batch normalization)https://github.com/martin-gorner/tensorflow-mnist-tutorial

6. 求AUC

import tensorflow as tf

a = tf.Variable([0.1, 0.5])
b = tf.Variable([0.2, 0.6])

auc = tf.contrib.metrics.streaming_auc(a, b)

sess = tf.Session()
sess.run(tf.initialize_all_variables())
sess.run(tf.initialize_local_variables()) # try commenting this line and you'll get the error
train_auc = sess.run(auc)

print(train_auc)

https://stackoverflow.com/questions/39435341/how-to-calculate-auc-with-tensorflow

 7. 可视化 （好）

Tensorflow的可视化工具Tensorboard的初步使用

https://blog.csdn.net/mao_feng/article/details/54731098

tensorflow + mnist + cnn + batch normalization

https://github.com/hwalsuklee/tensorflow-mnist-cnn

8. 过拟合情况及解决

dropout https://blog.csdn.net/smf0504/article/details/55254818

batch normalization https://blog.csdn.net/whitesilence/article/details/75667002

slim cnn 代码 https://www.2cto.com/kf/201706/649266.html

9.论文写法

https://blog.csdn.net/fengbingchun/article/details/50529500 cnn 结构