├── Arithmetic ├── MandelbrotSet.md ├── PartialDifferentialEquations.md └── WideDeepModel.md ├── BaseLearn ├── README.md ├── Supervisor │ ├── README.md │ └── SupervisorEx.py ├── install │ └── README.md ├── read_data │ ├── README.md │ ├── create_data.png │ ├── n1.csv │ ├── n2.csv │ └── read_data.py ├── share_variable │ ├── README.md │ └── share_variable.py ├── start │ ├── README.md │ └── start.py ├── tensorboard │ ├── README.md │ ├── SummayTest.py │ ├── TensorBoard.py │ ├── tensorboard_commandline.png │ ├── tensorboard_example.png │ └── test_1 │ │ ├── events.out.tfevents.1502120649.NIEBIN │ │ └── events.out.tfevents.1502121636.NIEBIN ├── tensorboard_image │ ├── README.md │ ├── SummayImage.py │ ├── start_person.jpg │ ├── sum_image.png │ └── test_2 │ │ ├── events.out.tfevents.1502182056.NIEBIN │ │ ├── events.out.tfevents.1502182108.NIEBIN │ │ ├── events.out.tfevents.1502182134.NIEBIN │ │ ├── events.out.tfevents.1502185571.NIEBIN │ │ ├── events.out.tfevents.1502185594.NIEBIN │ │ ├── events.out.tfevents.1502186495.NIEBIN │ │ └── events.out.tfevents.1502186504.NIEBIN ├── threads │ ├── README.md │ └── threadExample.py └── variable │ ├── README.md │ └── variable.py ├── LICENSE ├── PicClassify ├── Newer │ ├── README.md │ └── mnist_softmax.py ├── README.md ├── __init__.py └── mnist.py ├── README.md └── Resource ├── README.md ├── thread_queque.png └── 训练集.png /Arithmetic/MandelbrotSet.md: -------------------------------------------------------------------------------- 1 | # Mandelbrot Set 2 | 3 | 在此部分，介绍一个使用tensorflow可视化Mandelbrot set的例子。以这个例子表明，tensorflow可用于基础数学。以下内容来源于tensorflow教程部分[Mandelbrot Set](https://www.tensorflow.org/tutorials/mandelbrot)，对其进行了翻译和整理。 4 | 5 | 目录: 6 | 7 | * 什么是Mandelbrot set 8 | * 可视化Mandelbrot set 9 | 10 | ## 什么是Mandelbrot set（曼德博集合）？ 11 | 12 | Mandelbrot set是在[复平面](https://zh.wikipedia.org/wiki/%E5%A4%8D%E5%B9%B3%E9%9D%A2)组成[分形](https://zh.wikipedia.org/wiki/%E5%88%86%E5%BD%A2)的点的集合，以美国数学家Mandelbrot命名。 13 | 14 | Mandelbrot set是一个复数点c的集合，任取复平面上的一个点c，若满足：Z(n+1) = Z(n)^2 + c，Z(0)=0，当n不断增大时，|Z(n)|有界，则c属于Mandelbrot set。 15 | 16 | 在实际中，一般取一个较大的n值，迭代计算Z(n)是否有界（小于某个值），若有界则c属于Mandelbrot set。在可视化的时候，可选择在平面上c点的坐标以某种 颜色显示。 17 | 18 | ## 可视化Mandelbrot set 19 | 20 | ### import相关的库 21 | 22 | 导入仿真计算的库 23 | 24 | ```python 25 | import tensorflow as tf 26 | import numpy as np 27 | ``` 28 | 29 | 导入可视化的库 30 | ```python 31 | import PIL.Image 32 | from io import BytesIO 33 | from IPython.display import Image, display 34 | ``` 35 | 36 | ### 定义一个可视化函数 37 | 38 | ```python 39 | def DisplayFractal(a, fmt='jpeg'): 40 | a_cyclic = (6.28*a/20.0).reshape(list(a.shape)+[1]) 41 | img = np.concatenate([10+20*np.cos(a_cyclic), 42 | 30+50*np.sin(a_cyclic), 43 | 155-80*np.cos(a_cyclic)], 2) 44 | img[a==a.max()] = 0 45 | a = img 46 | a = np.uint8(np.clip(a, 0, 255)) 47 | f = BytesIO() 48 | PIL.Image.fromarray(a).save(f, fmt) 49 | display(Image(data=f.getvalue())) 50 | 51 | ``` 52 | 53 | ### 使用tensorflow进行一些数学计算 54 | 55 | ```python 56 | sess = tf.InteractiveSession() # 打开一个交互式的session，在这个session下进行计算 57 | 58 | Y, X = np.mgrid[-1.3:1.3:0.005, -2:1:0.005] # 创建复平面上点的横纵坐标 59 | Z = X+1j*Y # 复数Z 60 | 61 | xs = tf.constant(Z.astype(np.complex64)) # xs为tf的常量，其值为Z转成np.complex64类型后的值 62 | zs = tf.Variable(xs) # zs为tf中的变量，其初始值为xs 63 | ns = tf.Variable(tf.zeros_like(xs, tf.float32)) # ns为tf中的变量，与xs的shape相同，但值都为0，tf.float32类型。 64 | 65 | tf.global_variables_initializer().run() # 计算前先初始化所有变量 66 | 67 | zs_ = zs*zs + xs # 对应Mandelbrot set的迭代计算公式 68 | 69 | not_diverged = tf.abs(zs_) < 4 # 小于4时即收敛，与zs_的shape相同，但所有值为bool型 70 | 71 | step = tf.group( 72 | zs.assign(zs_), # 将zs_赋值给zs 73 | ns.assign_add(tf.cast(not_diverged, tf.float32)) # ns = ns+not_diverged 74 | ) 75 | 76 | for i in range(200): 77 | step.run() # 迭代200次 78 | 79 | ``` 80 | 81 | ### 可视化结果 82 | 83 | ```python 84 | DisplayFractal(ns.eval()) # 以上代码在ipython中运行时，并未出现可视化的图片。在jupyter notebook中，出现可视化的结果。 85 | ``` 86 | ![](https://www.tensorflow.org/images/mandelbrot_output.jpg) 87 | 88 | 89 | 参考文献： 90 | 91 | [1] [曼德布洛特集合简要概述](https://msdn.microsoft.com/zh-cn/library/jj635753(v=vs.85).aspx) 92 | 93 | [2] [tensorflow中的Mandelbrot set教程](https://www.tensorflow.org/tutorials/mandelbrot) 94 | 95 | 96 | 97 | 98 | -------------------------------------------------------------------------------- /Arithmetic/PartialDifferentialEquations.md: -------------------------------------------------------------------------------- 1 | # Partial Differential Equations（偏微分方程） 2 | 3 | TensorFlow不仅仅用于机器学习。在这部分，给出一个使用TensorFlow仿真偏微分方程的例子，模拟一些雨点落在一个方形池塘的水面上。 4 | 5 | 目录 6 | 7 | * 建立基础 8 | * 便于计算的函数 9 | * 定义概率密度函数 10 | * 运行仿真 11 | 12 | ## 建立基础 13 | 14 | 导入一些库 15 | ```python 16 | # 导入用于仿真的库 17 | import tensorflow as tf 18 | import numpy as np 19 | 20 | # 导入用于可视化的库 21 | import PIL.Image 22 | from io import BytesIO 23 | from IPython.display import clear_output, Image, display 24 | 25 | # 用于显示池塘水面状态的函数，水面状态以图像表示 26 | def DisplayArray(a, fmt='jpeg', rng=[0,1]): 27 | """Display an array as a picture.""" 28 | a = (a - rng[0])/float(rng[1] - rng[0])*255 29 | a = np.uint8(np.clip(a, 0, 255)) 30 | f = BytesIO() 31 | PIL.Image.fromarray(a).save(f, fmt) 32 | clear_output(wait = True) 33 | display(Image(data=f.getvalue())) 34 | 35 | # 打开一个交互式的session 36 | sess = tf.InteractiveSession() 37 | ``` 38 | 39 | ## 便于计算的函数 40 | 41 | ```python 42 | def make_kernel(a): 43 | """转换一个2维数组成一个卷积核""" 44 | a = np.asarray(a) 45 | a = a.reshape(list(a.shape) + [1,1]) 46 | return tf.constant(a, dtype=1) 47 | 48 | def simple_conv(x, k): 49 | """一个简单的2维卷积操作""" 50 | x = tf.expand_dims(tf.expand_dims(x, 0), -1) 51 | y = tf.nn.depthwise_conv2d(x, k, [1, 1, 1, 1], padding='SAME') 52 | return y[0, :, :, 0] 53 | 54 | def laplace(x): 55 | """计算一个数组的拉普拉斯变换""" 56 | laplace_k = make_kernel([[0.5, 1.0, 0.5], 57 | [1.0, -6., 1.0], 58 | [0.5, 1.0, 0.5]]) 59 | return simple_conv(x, laplace_k) 60 | ``` 61 | 62 | 其中，[tf.nn.depthwise_conv2d](https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d)的 63 | 图示为![](https://qiita-image-store.s3.amazonaws.com/0/25288/6f0d9da1-8f53-cf43-883a-fe55464376b8.png) 64 | 65 | ## 定义概率密度函数 66 | 67 | 假设池塘的尺寸为N*N，N=500。 68 | ```python 69 | # 池塘的初始状态，初始化为0 70 | u_init = np.zeros([N, N], dtype=np.float32) 71 | ut_init = np.zeros([N, N], dtype=np.float32) 72 | # 用几个随机点表示雨点，落在池塘里 73 | for n in range(40): 74 | a,b = np.random.randint(0, N, 2) 75 | u_init[a,b] = np.random.uniform() 76 | # 显示效果 77 | DisplayArray(u_init, rng=[-0.1, 0.1]) 78 | ``` 79 | 接下来，定义微分方程。 80 | ```python 81 | # 参数 82 | # eps -- time resolution 83 | # damping -- wave damping 84 | eps = tf.placeholder(tf.float32, shape=()) 85 | damping = tf.placeholder(tf.float32, shape=()) 86 | # 创建表示仿真状态的变量 87 | U = tf.Variable(u_init) 88 | Ut = tf.Variable(ut_init) 89 | # 离散化的概率密度函数的更新规则 90 | U_ = U + eps * Ut 91 | Ut_ = Ut + eps * (laplace(U) - damping * Ut) 92 | # 更新状态的操作 93 | step = tf.group( 94 | U.assign(U_), 95 | Ut.assign(Ut_)) 96 | ``` 97 | 98 | ## 运行仿真 99 | ```python 100 | # 初始化所有变量 101 | tf.global_variables_initializer().run() 102 | # 概率密度函数，运行1000次，每一次都输出一个池塘的表面状态结果 103 | for i in range(1000): 104 | # Step simulation 105 | step.run({eps: 0.03, damping: 0.04}) 106 | DisplayArray(U.eval(), rng=[-0.1, 0.1]) 107 | ``` 108 | 最终结果![](https://www.tensorflow.org/images/pde_output_2.jpg) 109 | 110 | 参考文献： 111 | [1] 112 | [tf中几种卷积的介绍](http://qiita.com/YusukeSuzuki@github/items/0764d15b9d0b97ec1e16) 113 | -------------------------------------------------------------------------------- /Arithmetic/WideDeepModel.md: -------------------------------------------------------------------------------- 1 | # Tensorflow Wide & Deep Learning Tutorial 2 | 3 | 此部分，简单介绍了一个使用tensorflow在一份数据集上训练wide & deep模型，并使用训练好的模型进行预测的例子。全部内容基于[tf官网例子](https://www.tensorflow.org/tutorials/wide_and_deep)进行了整理。tf官网上该例子的[源代码](https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/examples/learn/wide_n_deep_tutorial.py)，请结合源代码看此部分教程。 4 | 5 | ## 涉及到： 6 | * 数据集：[Census Income Data](https://archive.ics.uci.edu/ml/datasets/Census+Income) 7 | * 模型：Wide & Deep Model [论文](https://arxiv.org/abs/1606.07792)，[理论介绍](http://www.shuang0420.com/2017/03/13/%E8%AE%BA%E6%96%87%E7%AC%94%E8%AE%B0%20-%20Wide%20and%20Deep%20Learning%20for%20Recommender%20Systems/) 8 | * pandas（python的一个数据处理库，在这里用于读取数据集） 9 | * tempfile（python的一个临时文件相关的库，在这里用于创建临时文件，关闭后会自动删除） 10 | * [tf.contrib.learn.DNNLinearCombinedClassifier](https://www.tensorflow.org/api_docs/python/tf/contrib/learn/DNNLinearCombinedClassifier)（Wide & Deep Model在tensorflow中的实现） 11 | * [tf.contrib.layers.sparse_column_with_keys](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/sparse_column_with_keys)（tf中处理分类型特征的一种方式，当知道该特征中不同的特征值时使用） 12 | * [tf.contrib.layers.sparse_column_with_hash_bucket](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/sparse_column_with_hash_bucket)（tf中处理分类型特征的另一种方式，如果不知道该特征中不同的特征值时使用，但特征值必须时字符串或整型） 13 | * [tf.contrib.layers.real_valued_column](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/real_valued_column)（tf中处理连续型特征的一种方式） 14 | * [tf.contrib.layers.bucketized_column](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/bucketized_column)（tf中离散化连续型特征的一种方式） 15 | * [tf.contrib.layers.crossed_column](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/crossed_column)（tf中处理交叉特征的一种方式） 16 | * [tf.contrib.layers.embedding_column](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/embedding_column)（tf中embedding稀疏特征，并feed给DNN） 17 | 18 | ## 目录： 19 | * Census Income Data介绍和下载 20 | * Wide & Deep Model简介 21 | * 数据预处理 22 | * 模型的训练和评估 23 | 24 | ！！！请注意！！！为避免对有些专有名词翻译的不准确，在这里将不翻译。 25 | 26 | ## 1、 Census Income Data介绍和下载 27 | 28 | ### 1.1、 Census Income Data介绍 29 | 30 | 数据来源于美国1994年人口调查数据库。[数据集介绍](https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.names)。 31 | 32 | 样本个数：train=32561, test=16281。 33 | 类别：>50K, <=50K。 34 | 特征个数：14个，其中： 35 | 分类型特征：CATEGORICAL_COLUMNS = ["workclass", "education", "marital_status", "occupation", 36 | "relationship", "race", "gender", "native_country"] 37 | 连续型特征：CONTINUOUS_COLUMNS = ["age", "education_num", "capital_gain", "capital_loss", 38 | "hours_per_week"] 39 | 40 | ### 1.2、Census Income Data下载 41 | 42 | 1） 手动下载，[数据集地址](https://archive.ics.uci.edu/ml/machine-learning-databases/adult/) 43 | 44 | 2） 使用[源代码](https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/examples/learn/wide_n_deep_tutorial.py)中的maybe_download函数下载 45 | 46 | ### 2、Wide & Deep Model简介 47 | 48 | ![Wide & Deep Model](https://www.tensorflow.org/images/wide_n_deep.svg) 49 | 50 | 上图左为Wide Model；上图右为Deep Model；上图中为Wide & Deep Model，该模型由两个模型组成：Wide Model（线性模型），Deep Model（DNN模型）。在tensorflow中有该模型的实现：[tf.contrib.learn.DNNLinearCombinedClassifier](https://www.tensorflow.org/api_docs/python/tf/contrib/learn/DNNLinearCombinedClassifier)（模型的理论介绍，可从本篇开头介绍的链接找到）。DNNLinearCombinedClassifier函数的参数有linear_feature_columns，dnn_feature_columns，dnn_hidden_units等。其中linear_feature_columns为模型的wide part的输入，包括sparse features和sparse crossed features（交叉特征）；dnn_feature_columns为模型的deep part的输入，包括连续型特征，每一个分类型特征的embedding。 51 | 52 | 53 | 54 | ### 3、数据预处理 55 | 56 | 1） 原数据集中含有缺失值，而模型的输入不能含有缺失值，需去除含有缺失值的样本。在[源代码](https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/examples/learn/wide_n_deep_tutorial.py)的train_and_eval函数中有相应处理，利用了pandas的dropna函数。 57 | 58 | 我手动下载了数据集，发现其中的缺失值用'?'代替了。我的处理方案时，使用pd.read_csv读取数据集，返回dataframe对象，设为df。 59 | 60 | l = df.apply(lambda x:'?' in x.values, axis=1) # 如果某一行含有'?'，则为True，否则为False。 61 | df1 = df.drop(df.loc[l].index) # 利用得到的l来索引df，得到含有'?'的行，然后去除这些行，得到df1。 62 | 63 | 2）对于分类型特征，由于其值本身没有任何意义，一般使用OneHotEncoder来处理。在tensorflow中有相应的函数[tf.contrib.layers.sparse_column_with_keys](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/sparse_column_with_keys)，[tf.contrib.layers.sparse_column_with_hash_bucket](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/sparse_column_with_hash_bucket)，将分类型特征转成sparse features，这里并没有执行，只是先进行了定义，用于feed给DNN，下面类似的几个函数也是进行类似的定义。 64 | 65 | 将若干个分类型特征组成交叉特征的函数为[tf.contrib.layers.crossed_column](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/crossed_column)，特征工程的一种操作。 66 | 67 | 将两个类别转成0和1表示的方法，见源代码中的177和178行。 68 | 69 | 对于有些特征值很多的特征，虽然可将其转成Sparse features，但维度高，还需使用[tf.contrib.layers.embedding_column](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/embedding_column)对其降维转成embedding vecotr，才能作为wide part的输入。对于降维后其维度的确定，可参考 70 | [官网教程的The Deep Model: Neural Network with Embeddings部分](https://www.tensorflow.org/tutorials/wide_and_deep)，设置为log2(n)或者k*(n)**(1/4)，n为该特征不同的特征值的个数，k为一个小的常数，通常小于10。 71 | 72 | 3）对于连续型特征，使用了函数[tf.contrib.layers.real_valued_column](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/real_valued_column)来处理。 73 | 74 | 4）[源代码](https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/examples/learn/wide_n_deep_tutorial.py)的input_fn函数，将传入pandas读取的dataframe对象df的每一列，转成tensorflow中的tensor，并创建特征名与其tensor的Python字典。 75 | 76 | 对于连续型特征，使用tf.constant(df[特征名].values)转成一维的tensor； 77 | 78 | 对于分类型特征，使用[tf.SparseTensor](https://www.tensorflow.org/api_docs/python/tf/SparseTensor)进行转换。该函数中最重要的三个参数为indice（二维数组，表示dense feature中非0值的下标），values（一维数组，表示feature的非0值），dense_shape（一维数组，表示dense feature的shape）。 79 | 80 | （个人观点：从源代码中tf.SparseTensor的使用来看，对于传递给该函数的每一个分类型特征，indice=[..., [i,0], ...]； value=一个dataframe列的所有值，一维数组；dense_shape=[该列的长度，1]。如果将其转成dense tensor，结果与该列的唯一区别就是维度从一维，变成了二维的[该列的长度，1]。这样的话，并没有达到sparse的特点，请高人指点。） 81 | 82 | 83 | ### 4、 模型的训练和评估 84 | 85 | 在[源代码](https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/examples/learn/wide_n_deep_tutorial.py)中，build_estimator函数用于创建wide& deep model，返回模型的对象m。 86 | 87 | 然后调用模型的[fit方法](https://www.tensorflow.org/api_docs/python/tf/contrib/learn/DNNLinearCombinedClassifier)，传入数据、迭代次数等参数进行训练。 88 | 89 | 调用模型的[evaluate方法](https://www.tensorflow.org/api_docs/python/tf/contrib/learn/DNNLinearCombinedClassifier)进行预测。 90 | -------------------------------------------------------------------------------- /BaseLearn/README.md: -------------------------------------------------------------------------------- 1 | [安装](./install) 2 | [快速入门](./start) 3 | [变量相关的](./variable) 4 | -------------------------------------------------------------------------------- /BaseLearn/Supervisor/README.md: -------------------------------------------------------------------------------- 1 | # 训练助手 2 | > 当需要做以日计的训练数据 有如下需求 3 | - 处理关机和异常的清除 4 | - 可以从关机或异常中重新运行 5 | - 可以TensorBoard中被监控 6 | 7 | 关键类：[tf.train.Supervisor](https://www.tensorflow.org/api_docs/python/tf/train/Supervisor) 8 | 9 | ```python 10 | def runSum(): 11 | a=tf.constant(12,dtype=tf.int8) 12 | b=tf.constant(10,dtype=tf.int8) 13 | sv=tf.train.Supervisor(logdir="./test1") 14 | with sv.managed_session() as sess: 15 | for i in range(10): 16 | if sv.should_stop(): 17 | return 18 | print(sess.run([a,b])) 19 | ''' 20 | [12, 10] 21 | [12, 10] 22 | [12, 10] 23 | [12, 10] 24 | [12, 10] 25 | [12, 10] 26 | [12, 10] 27 | [12, 10] 28 | [12, 10] 29 | [12, 10] 30 | ''' 31 | ``` 32 | [实例代码](./SupervisorEx.py) 33 | - 未完待续 34 | -------------------------------------------------------------------------------- /BaseLearn/Supervisor/SupervisorEx.py: -------------------------------------------------------------------------------- 1 | # user/bin 2 | import tensorflow as tf 3 | import threading 4 | ''' 5 | this is test for how to use the Supervisor 6 | ''' 7 | def runSum(): 8 | a=tf.constant(12,dtype=tf.int8) 9 | b=tf.constant(10,dtype=tf.int8) 10 | sv=tf.train.Supervisor(logdir="./test1") 11 | with sv.managed_session() as sess: 12 | for i in range(10): 13 | if sv.should_stop(): 14 | return 15 | print(sess.run([a,b])) 16 | 17 | runSum() 18 | -------------------------------------------------------------------------------- /BaseLearn/install/README.md: -------------------------------------------------------------------------------- 1 | ## 安装window-cpu版本   2 | 1. 安装python3.5.2 3 | > 一定要是这个版本，如果是其他版本会安装不上 4 | 一配置好环境变量 5 | PYTHON_PATH=E:\Tools\Python35\Scripts;E:\Tools\Python35 6 | 7 | 2. 安装tensorflow 8 | 输入命令：pip3 install --upgrade tensorflow 9 | 3. 验证 10 | 命令：import tensorflow as tf 11 | - a.No module named "pywrap_tensorflow" 12 | 解决方案：下载https://www.microsoft.com/en-us/download/details.aspx?id=53587 13 | - b.NameError: name 'core' is not defined 14 | - c.NameError: name 'python' is not defined 15 | #### bc重启命令好久好了!! 16 | 17 | ## Mac 18 | 等待你的完成 19 | ## Linux 20 | 等待你的参与 21 | -------------------------------------------------------------------------------- /BaseLearn/read_data/README.md: -------------------------------------------------------------------------------- 1 | # 读取数据 2 | > 以读取csv文件(可以用excel进行转换 可以作为普通文本打开)为例 如果是普通文件 可以直接用python的接口读取 3 | 4 | - ## 绑定文件 5 | > 将文件放在队列中 创建reader 绑定到队列中 进行操作 6 | > 关键类 TextLineReader 7 | ``` python 8 | filename_queue = tf.train.string_input_producer(["n1.csv", "n2.csv"])#将文件扔进管道里面 9 | reader = tf.TextLineReader() 10 | key, value = reader.read(filename_queue)#读取文件的 , 11 | print("key=",key)# 打印 key= Tensor("ReaderReadV2:0", shape=(), dtype=string) 12 | print("value=",value)#value= Tensor("ReaderReadV2:1", shape=(), dtype=string) 13 | ``` 14 | 15 | - ## 特殊格式解析 16 | > 这一步 主要是根据文件格式 选择对应的解析器 对文件进行解析 并根据内容进行特殊处理 17 | > 关键词 tf.decode_csv 18 | ```python 19 | record_defaults = [[1], [1], [1], [1], [1]] 20 | col1, col2, col3, col4, col5 = tf.decode_csv( 21 | value, record_defaults=record_defaults)# 进行解析 22 | print("c1-5=",col1," ",col2," ",col3," ",col4," ",col5) 23 | features = tf.stack([col1, col2, col3, col4]) 24 | print("f=",features) 25 | ``` 26 | 27 | - ## 读取内容 28 | > 终于到了最关键的一步 前面都是准备工作 ，这才是真正的数据读取 用法也很简单 29 | > 关键字:coordinater,线程 不懂的请参照[队列和线程](https://github.com/TFStudents/Tensorflow/tree/master/BaseLearn/threads) 30 | ```python 31 | with tf.Session() as sess: 32 | # Start populating the filename queue. 33 | coord = tf.train.Coordinator() 34 | threads = tf.train.start_queue_runners(coord=coord) 35 | for i in range(5): 36 | # Retrieve a single instance: 37 | example, label = sess.run([features, col5]) 38 | print(example) 39 | print(label) 40 | coord.request_stop() 41 | coord.join(threads) 42 | ``` 43 | 44 | ## 资源 45 | [参考地址](https://www.tensorflow.org/programmers_guide/reading_data) 46 | [n1.csv](n1.csv) 47 | [n2.csv](n2.csv) 48 | [本实例代码](read_data.py) 49 | ![流程图](create_data.png) 50 | -------------------------------------------------------------------------------- /BaseLearn/read_data/create_data.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/read_data/create_data.png -------------------------------------------------------------------------------- /BaseLearn/read_data/n1.csv: -------------------------------------------------------------------------------- 1 | 1,2,3,4,5 2 | -------------------------------------------------------------------------------- /BaseLearn/read_data/n2.csv: -------------------------------------------------------------------------------- 1 | 101,102,103,104,105 2 | 201,203,204,202,205 3 | -------------------------------------------------------------------------------- /BaseLearn/read_data/read_data.py: -------------------------------------------------------------------------------- 1 | # user/bin 2 | 3 | import tensorflow as tf 4 | import threading 5 | import time 6 | """ 7 | n2.csv content 8 | 101,102,103,104,105 9 | 201,203,204,202,205 10 | 11 | n1.csv content 12 | 1,2,3,4,5 13 | """ 14 | # 1 read file 15 | filename_queue = tf.train.string_input_producer(["n1.csv", "n2.csv"]) 16 | sess=tf.Session() 17 | reader = tf.TextLineReader() 18 | key, value = reader.read(filename_queue) 19 | print("key=",key) 20 | print("value=",value) 21 | # Default values, in case of empty columns. Also specifies the type of the 22 | # decoded result 23 | 24 | # 2 decode file 25 | record_defaults = [[1], [1], [1], [1], [1]] 26 | col1, col2, col3, col4, col5 = tf.decode_csv( 27 | value, record_defaults=record_defaults) 28 | print("c1-5=",col1," ",col2," ",col3," ",col4," ",col5) 29 | features = tf.stack([col1, col2, col3, col4]) 30 | print("f=",features) 31 | # 3 get the file content 32 | with tf.Session() as sess: 33 | # Start populating the filename queue. 34 | coord = tf.train.Coordinator() 35 | threads = tf.train.start_queue_runners(coord=coord) 36 | for i in range(5): 37 | # Retrieve a single instance: 38 | example, label = sess.run([features, col5]) 39 | print(example) 40 | print(label) 41 | coord.request_stop() 42 | coord.join(threads) -------------------------------------------------------------------------------- /BaseLearn/share_variable/README.md: -------------------------------------------------------------------------------- 1 | # 共享变量 2 | [share_variable.py](./share_variable.py) 3 | ## 存在问题 4 | - 代码块 A 5 | ``` 6 | sess=tf.Session() 7 | def printVariable(): 8 | v1=tf.Variable(1,name="v1") 9 | v2=tf.Variable(1,name="v2") 10 | init=tf.global_variables_initializer()#变量的初始化 初始化当前module所有的变量 11 | sess.run(init) 12 | return v1+v2 13 | 14 | a1=printVariable()#两个变量 15 | b1=printVariable()#两个变量 总4个变量 16 | print(sess.run(a1)) 17 | print(sess.run(b1)) 18 | ``` 19 | 当我们运行代码块A时,方法内会产生4个变量，当运算大时，会产生过多的内存消耗，那怎么解决这个问题呢？ 20 | 有如下方案 21 | - 代码块B 22 | ``` 23 | v_dict={ 24 | "v1":tf.Variable(1,name="v1"), 25 | "v2":tf.Variable(1,name="v2") 26 | } 27 | def printVariable(dicts): 28 | v1=dicts["v1"] 29 | v2=dicts["v2"] 30 | return v1.initialized_value()+v2.initialized_value() 31 | a2=printVariable(v_dict)#两个变量 32 | b2=printVariable(v_dict)#两个变量 总2两个 33 | print(sess.run(a2)) 34 | print(sess.run(b2)) 35 | ``` 36 | 当我们执行两遍printVariable方法时，确实解决了这个问题，只有两个变量，单存在一些问题 37 | 1. 当创建变量时，必须指定一些如类型，形状等参数 38 | 2. 当代码改变时，可能创建的变量会更多 39 | 40 | tensorflow提供了一种更方便的方式提供相应的功能，直接看代码 41 | - 代码块C 42 | ``` 43 | def printVariable2(): 44 | v1=tf.get_variable("v1",[1],initializer=tf.constant_initializer(1)) 45 | print("v1.name="+v1.name) 46 | v2=tf.get_variable("v2",[1],initializer=tf.constant_initializer(2)) 47 | print("v2.name="+v2.name) 48 | return v1+v2 49 | with tf.variable_scope("boo") as boo: 50 | printVariable2() 51 | print("----------boo--1-----------") 52 | boo.reuse_variables() 53 | printVariable2() 54 | print("----------boo--2-----------") 55 | printVariable2() 56 | print("----------boo--3-----------") 57 | printVariable2() 58 | print("----------boo--4-----------") 59 | printVariable2() 60 | ''' 61 | 打印的信息 62 | v1.name=boo/v1:0 63 | v2.name=boo/v2:0 64 | ----------boo--1----------- 65 | v1.name=boo/v1:0 66 | v2.name=boo/v2:0 67 | ----------boo--2----------- 68 | v1.name=boo/v1:0 69 | v2.name=boo/v2:0 70 | ----------boo--3----------- 71 | v1.name=boo/v1:0 72 | v2.name=boo/v2:0 73 | ----------boo--4----------- 74 | v1.name=boo/v1:0 75 | v2.name=boo/v2:0 76 | 77 | ''' 78 | ``` 79 | 上面虽然调用了好几次，但是变量是共享的，所以只产生了2个变量,而tf就是这么进行变量共享的，先看两个关键函数 80 | ## tf.get_variable 81 | [tf.get_variable](https://www.tensorflow.org/api_docs/python/tf/get_variable) 82 | - 参数列表 83 | ``` 84 | get_variable( 85 | name, 86 | shape=None, 87 | dtype=None, 88 | initializer=None, 89 | regularizer=None, 90 | trainable=True, 91 | collections=None, 92 | caching_device=None, 93 | partitioner=None, 94 | validate_shape=True, 95 | use_resource=None, 96 | custom_getter=None 97 | ) 98 | ``` 99 | 传入值时，除name外shape也是必须传的 100 | ``` 101 | v=tf.get_variable("v1",shape=[1]，initializer=tf.constant_initializer(1)) #这样一个变量就创建了 102 | ``` 103 | - 它是怎么工作的？ 104 | 1. 名字： 传入相应参数后，v将会被创建，它的名字就是scope_name+variable_name 105 | 在代码块C中,scope_name="boo",variable_name="v1"或"v2" 则v1的全名就是"boo/v1:0" 参考打印信息 106 | 2. 重用：要确保tf.get_variable_scope().reuse == True ,调用就scope.reuse_variables()就可以了 但不能直接设置为false 107 | 108 | ## tf.variable_scope 109 | [tf.variable_scope](https://www.tensorflow.org/api_docs/python/tf/variable_scope) 110 | - 参数列表 111 | ``` 112 | variable_scope( 113 | name_or_scope, 114 | default_name=None, 115 | values=None, 116 | initializer=None, 117 | regularizer=None, 118 | caching_device=None, 119 | partitioner=None, 120 | custom_getter=None, 121 | reuse=None, 122 | dtype=None, 123 | use_resource=None 124 | ) 125 | ``` 126 | 一般只传name值就可以了，作用就是为tf.get_variable做名字区分以及重用控制，当然还可以控制initializer 127 | 差不多这些内容就够用了 ，具体细节点击链接看看[本文参考地址](https://www.tensorflow.org/programmers_guide/variable_scope) 128 | 129 | ## end 130 | 131 | -------------------------------------------------------------------------------- /BaseLearn/share_variable/share_variable.py: -------------------------------------------------------------------------------- 1 | #user/bin 2 | import tensorflow as tf 3 | sess=tf.Session() 4 | print("----------0-------------") 5 | def printVariable(): 6 | v1=tf.Variable(1,name="v1") 7 | v2=tf.Variable(1,name="v2") 8 | init=tf.global_variables_initializer()#变量的初始化 初始化当前module所有的变量 9 | sess.run(init) 10 | return v1+v2 11 | 12 | a1=printVariable()#两个变量 13 | b1=printVariable()#两个变量 总4个变量 14 | print(sess.run(a1)) 15 | print(sess.run(b1)) 16 | print("----------1-------------") 17 | v_dict={ 18 | "v1":tf.Variable(1,name="v1"), 19 | "v2":tf.Variable(1,name="v2") 20 | } 21 | def printVariable(dicts): 22 | v1=dicts["v1"] 23 | v2=dicts["v2"] 24 | return v1.initialized_value()+v2.initialized_value() 25 | a2=printVariable(v_dict)#两个变量 26 | b2=printVariable(v_dict)#两个变量 总2两个 27 | print(sess.run(a2)) 28 | print(sess.run(b2)) 29 | 30 | print("----------2-------------") 31 | 32 | def printVariable2(): 33 | v1=tf.get_variable("v1",[1],initializer=tf.constant_initializer(1)) 34 | print("v1.name="+v1.name) 35 | v2=tf.get_variable("v2",[1],initializer=tf.constant_initializer(2)) 36 | print("v2.name="+v2.name) 37 | return v1+v2 38 | with tf.variable_scope("boo") as boo: 39 | printVariable2() 40 | print("----------boo--1-----------") 41 | boo.reuse_variables() 42 | printVariable2() 43 | print("----------boo--2-----------") 44 | printVariable2() 45 | print("----------boo--3-----------") 46 | printVariable2() 47 | print("----------boo--4-----------") 48 | printVariable2() 49 | 50 | #------------variable_scope-------------------------- 51 | def onTestScope(): 52 | with tf.variable_scope("boo") as boo: 53 | v=tf.get_variable("v_1",[20],initializer=tf.constant_initializer(2)) 54 | print("v.name="+v.name) 55 | print("boo.name="+boo.name) 56 | with tf.variable_scope(boo): 57 | v2=tf.get_variable("v_2",[20],initializer=tf.constant_initializer(3)) 58 | v1=v+v2 59 | print("v2.name="+v2.name) 60 | print("v1.name="+v1.name) 61 | return v1 62 | with tf.variable_scope("test") as sc: 63 | onTestScope() 64 | sc.reuse_variables() 65 | onTestScope() 66 | 67 | # with tf.variable_scope("ABC",initializer=tf.constant_initializer(233)): 68 | # v=tf.get_variable("v3",[1]) 69 | # with tf.Session() as sess: 70 | # print(sess.run(v.variable_value())) 71 | 72 | 73 | 74 | -------------------------------------------------------------------------------- /BaseLearn/start/README.md: -------------------------------------------------------------------------------- 1 | # 快速入门 2 | [源码-start.py](./start.py) 3 | ```python 4 | #首先引入tensorflow 代码块是同一.py中 5 | import tensorflow as tf 6 | ``` 7 | ## 张量 8 | > 符号 Tensor 9 | 10 | - 阶 rank 11 | > 类似于维度 12 | > 什么是张量？张量相当于表,又有点类似于树 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 | 48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 |
阶数维度实际意义举例
00维张量标量2
11维张量向量[1,2]
22维张量矩阵[[123,33],[232,21]]
33维张量2维张量的向量[[1,2],[123,33]],[[3,5],[7,9]]]
44维张量3维张量的向量~~~
55维张量4维张量的向量~~~
nn维张量n-1维张量的向量~~~
63 | 64 | 65 | 66 | - 形状 Shape 67 | > 就是张量形状描述 ，长度是张量的阶数 68 | 69 | 70 | 71 | 72 | 73 | 74 | 75 | 76 | 77 | 78 | 79 | 80 | 81 | 82 | 83 | 84 | 85 | 86 | 87 | 88 | 89 |
阶shape举例
0[]shape[]
1[D0]#D0是数据集shape[233]#D0的个数
n[D0,....,Dn-1]#D0是数据集shape[D0,....,Dn-1]
90 | 91 | - 数据类型 type 92 | > 支持所有python 支持的基础数据类型 93 | 94 | - 流图Graph 95 | > 由一系列节点(node)汇聚而成的tf操作(option) 96 | 97 | ## 常量constant 98 | 方法[tf.constant](https://www.tensorflow.org/api_docs/python/tf/constant) 99 | 100 | - 构造函数 101 | ```python 102 | constant( 103 | value, 104 | dtype=None, 105 | shape=None, 106 | name='Const', 107 | verify_shape=False 108 | ) 109 | ``` 110 | 111 | - 举例 112 | ```python 113 | node1= tf.constant([True,True]) 114 | node2= tf.constant(12) 115 | node3=tf.constant("hello") 116 | node4=tf.constant("world") 117 | print(node1)#打印Tensor("Const:0", shape=(2,), dtype=bool) 118 | print(node2)#打印Tensor("Const_1:0", shape=(), dtype=int32) 119 | print(node3)#打印Tensor("Const_2:0", shape=(), dtype=string) 120 | sess=tf.Session()#获取会话 121 | print(sess.run(node1))#打印[True,True] 122 | print(sess.run(node2))#打印12 123 | print(sess.run(node3))#打印b'hello' 124 | print(sess.run(node4))#打印b'world' 125 | #要想获取平常的值如12 必须通过sess.run()进行，不然输入还是tensor格式 126 | ``` 127 | 128 | ## 加法 129 | 方法[tf.add](https://www.tensorflow.org/api_docs/python/tf/add) 130 | [更多数学运算](https://www.tensorflow.org/api_guides/python/math_ops#Arithmetic_Operators) 131 | 132 | - 构造函数 133 | ```python 134 | add( 135 | x, 136 | y, 137 | name=None 138 | ) 139 | ``` 140 | 141 | - 举例 142 | ```python 143 | node5=tf.constant(-1) 144 | n_add=tf.add(node2,node5) 145 | print(sess.run(n_add))#打印11 146 | node6=tf.constant(-2) 147 | n_add2=node5+node6 148 | print(sess.run(n_add2))#打印-3 说明两种形式的加法也可以的 149 | ``` 150 | 151 | ## 占位符 152 | 方法[tf.placeholder](https://www.tensorflow.org/api_docs/python/tf/placeholder) 153 | 相当于数学函数的右侧的变量 154 | 构造函数 155 | ```python 156 | placeholder( 157 | dtype, 158 | shape=None, 159 | name=None 160 | ) 161 | ``` 162 | 例子 163 | ```python 164 | p1=tf.placeholder(tf.float32) 165 | p2=tf.placeholder(tf.float32) 166 | n_add3=p1+p2 167 | a=sess.run(n_add3,{p1:23,p2:32})#用法正确 168 | #a=(n_add3,{p1:23,p2:32})错误 只能和sess一起使用 169 | print(a)#55.0 170 | print(sess.run(n_add3,{p1:23,p2:32}))#用法正确 打印55.0 171 | ``` 172 | 173 | ## 变量 174 | [tf.Variable](https://www.tensorflow.org/api_docs/python/tf/Variable) 175 | 176 | - 构造函数 177 | ```python 178 | __init__( 179 | initial_value=None, 180 | trainable=True, 181 | collections=None, 182 | validate_shape=True, 183 | caching_device=None, 184 | name=None, 185 | variable_def=None, 186 | dtype=None, 187 | expected_shape=None, 188 | import_scope=None 189 | ) 190 | ``` 191 | 192 | - 实例 193 | ```python 194 | v1=tf.Variable(12) 195 | v2=tf.Variable(23)#定义一个变量 196 | init = tf.global_variables_initializer()#进去全部变量的初始化 197 | sess.run(init) 198 | x=tf.placeholder(tf.int32) 199 | a_line=v1*x+v2#使用和常量类似 200 | v3=sess.run(a_line,{x:[1,2]}) 201 | print(v3)#[35 47] 202 | #重新赋值 203 | v1=tf.assign(v1,-2) 204 | v2=tf.assign(v2,3) 205 | add4=v1+v2 206 | v4=sess.run(add4) 207 | print(v4)#1 208 | 209 | ``` 210 | ## END 211 | -------------------------------------------------------------------------------- /BaseLearn/start/start.py: -------------------------------------------------------------------------------- 1 | #user/bin 2 | import tensorflow as tf 3 | #---常数---- 4 | node1= tf.constant([True,True]) 5 | node2= tf.constant(12) 6 | node3=tf.constant("hello") 7 | node4=tf.constant("world") 8 | print(node1)#打印Tensor("Const:0", shape=(2,), dtype=bool) 9 | print(node2)#打印Tensor("Const_1:0", shape=(), dtype=int32) 10 | print(node3)#打印Tensor("Const_2:0", shape=(), dtype=string) 11 | #---会话--- 12 | sess=tf.Session()#获取会话 13 | print(sess.run(node1))#打印[True,True] 14 | print(sess.run(node2))#打印12 15 | print(sess.run(node3))#打印b'hello' 16 | print(sess.run(node4))#打印b'world' 17 | #---加法---- 18 | node5=tf.constant(-1) 19 | n_add=tf.add(node2,node5) 20 | print(sess.run(n_add))#打印11 21 | node6=tf.constant(-2) 22 | n_add2=node5+node6 23 | print(sess.run(n_add2))#打印-3 说明两种形式的加法也可以的 24 | #---占位符--- 25 | p1=tf.placeholder(tf.float32) 26 | p2=tf.placeholder(tf.float32) 27 | n_add3=p1+p2 28 | a=sess.run(n_add3,{p1:23,p2:32})#用法正确 赋值以json的格式 进行赋值 29 | #a=(n_add3,{p1:23,p2:32})错误 只能和sess一起使用 30 | print(a)#55.0 31 | print(sess.run(n_add3,{p1:23,p2:32}))#用法正确 打印55.0 32 | #---变量--- 33 | v1=tf.Variable(12) 34 | v2=tf.Variable(23)#定义一个变量 35 | init = tf.global_variables_initializer()#进去全部变量的初始化 36 | sess.run(init) 37 | x=tf.placeholder(tf.int32) 38 | a_line=v1*x+v2#使用和常量类似 39 | v3=sess.run(a_line,{x:[1,2]}) 40 | print(v3)#[35 47] 41 | #重新赋值 42 | v1=tf.assign(v1,-2) 43 | v2=tf.assign(v2,3) 44 | add4=v1+v2 45 | v4=sess.run(add4) 46 | print(v4)#1 47 | -------------------------------------------------------------------------------- /BaseLearn/tensorboard/README.md: -------------------------------------------------------------------------------- 1 | [源码TensorBoard.py](./TensorBoard.py) 2 | # 如何使用TensorBoard 3 | >说明：tensorBoard是Tensorflow的可视化工具 4 | # 使用方法 5 | 1. 构建图并保存到文件 6 | ```python 7 | import tensorflow as tf 8 | node1=tf.constant(2,tf.float32)#定义节点，值为2 9 | node2=tf.constant(3,tf.float32)#定义节点，值为3 10 | node=tf.add(node1,node2) 11 | print(node1) 12 | with tf.Session() as sess: 13 | print(sess.run(node)) 14 | tf.summary.merge_all 15 | tf.global_variables_initializer().run() 16 | writer = tf.summary.FileWriter("test_1", sess.graph) 17 | 18 | ``` 19 | [生成文件](./test_1) 20 | 2. 显示 21 | > 利用1生成的文件，将视图关系显示出来，在命令行中输入 22 | ``` 23 | tensorboard --logdir="./test_1" #这是文件的路径 24 | ``` 25 | [运行时可能错误 由于 html5lib 版本问题，可以在anaconda中进行降版本解决此问题](https://github.com/tensorflow/tensorboard/pull/589) 26 | 27 | 3. 在浏览器中打开 28 | > 在浏览器中输入：http://192.168.56.1:6006 ，这是自己的电脑IP：6006是端口号     29 | - 而且在步骤2中有提示信息 30 | ![命令行](./tensorboard_commandline.png) 31 | - 效果图 32 | ![效果图](./tensorboard_example.png) 33 | -------------------------------------------------------------------------------- /BaseLearn/tensorboard/SummayTest.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard/SummayTest.py -------------------------------------------------------------------------------- /BaseLearn/tensorboard/TensorBoard.py: -------------------------------------------------------------------------------- 1 | #python 2 | import tensorflow as tf 3 | node1=tf.constant(2,tf.float32)#定义节点，值为2 4 | node2=tf.constant(3,tf.float32)#定义节点，值为3 5 | node=tf.add(node1,node2) 6 | print(node1) 7 | with tf.Session() as sess: 8 | print(sess.run(node)) 9 | tf.global_variables_initializer().run() 10 | writer = tf.summary.FileWriter("test_1", sess.graph) 11 | -------------------------------------------------------------------------------- /BaseLearn/tensorboard/tensorboard_commandline.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard/tensorboard_commandline.png -------------------------------------------------------------------------------- /BaseLearn/tensorboard/tensorboard_example.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard/tensorboard_example.png -------------------------------------------------------------------------------- /BaseLearn/tensorboard/test_1/events.out.tfevents.1502120649.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard/test_1/events.out.tfevents.1502120649.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard/test_1/events.out.tfevents.1502121636.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard/test_1/events.out.tfevents.1502121636.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/README.md: -------------------------------------------------------------------------------- 1 | # Tensorboard显示图片 2 | 基本使用我们会了，那现在来看看怎么使用图片吧。 3 | [实例代码](./SummaryImage.py) 4 | 完全不会使用Tensorboard-[请点击这里](../tensorboard) 5 | ![图片资源](./start_person.jpg) 6 | ### 1. 读取图片 7 | 8 | ```python 9 | image_filename = "start_person.jpg"# 图片地址，这里是放在当前文件夹的 10 | # 获取图片数据 11 | file = open(image_filename, 'rb') 12 | data = file.read() 13 | file.close() 14 | ``` 15 | ### 2. 解析图片 16 | 解析图片，并转换为Tensor,这一步很关键，注意这里加一维 17 | ```python 18 | file_img = tf.image.decode_jpeg(data,channels=3) #Tensor("DecodeJpeg:0", shape=(?, ?, 3), dtype=uint8) 19 | file_img=tf.expand_dims(file_img,0)# Tensor("ExpandDims:0", shape=(1, ?, ?, 3), dtype=uint8) 20 | ``` 21 | ### 3. 将图片写入文件 22 | ```python 23 | with tf.Session() as sess: 24 | init_op = tf.global_variables_initializer() 25 | sess.run(init_op) 26 | write = tf.summary.FileWriter("./test_2", graph=sess.graph) 27 | img = tf.summary.image("./", tensor=file_img) 28 | img_string = sess.run(img) 29 | write.add_summary(img_string) 30 | ``` 31 | [生成文件](./test_2) 32 | ### 4. 在Tensorboard中显示 33 | 在命令行，输入如下命令，在浏览器中输入类http://192.168.56.1:6006 类似   地址进行查看 34 | [参考](../tensorboard) 35 | ``` 36 | tensorboard --logdir="./test_2" 37 | ``` 38 | ### 5. 结果 39 | ![结果](./sum_image.png) 40 | -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/SummayImage.py: -------------------------------------------------------------------------------- 1 | #user/bin 2 | import tensorflow as tf 3 | 4 | def imageTest(): 5 | image_filename = "start_person.jpg"# 智能星人的图标 6 | # 获取图片数据 7 | file = open(image_filename, 'rb') 8 | data = file.read() 9 | file.close() 10 | # --- 11 | file_img = tf.image.decode_jpeg(data,channels=3) 12 | print(file_img) 13 | file_img=tf.expand_dims(file_img,0) 14 | print(file_img) 15 | with tf.Session() as sess: 16 | write = tf.summary.FileWriter("./test_2", graph=sess.graph) 17 | img = tf.summary.image("./", tensor=file_img) 18 | img_string = sess.run(img) 19 | write.add_summary(img_string) 20 | imageTest() -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/start_person.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/start_person.jpg -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/sum_image.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/sum_image.png -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502182056.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502182056.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502182108.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502182108.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502182134.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502182134.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502185571.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502185571.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502185594.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502185594.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502186495.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502186495.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502186504.NIEBIN: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/BaseLearn/tensorboard_image/test_2/events.out.tfevents.1502186504.NIEBIN -------------------------------------------------------------------------------- /BaseLearn/threads/README.md: -------------------------------------------------------------------------------- 1 | # 队列和线程 2 | ## 队列概述 3 | - 多个线程准备实例并放在队列中 4 | - 训练线程 从队列中取出训练操作op 5 | - Session 是多线程的，所以多线程可以使用同一个Session 6 | - 所有的线程都可以一起被关闭 并且当队列停止时要合理关闭 7 | 8 | 设计两个类 一起使用 9 | ## tf.train.Coordinator 10 | > 协作者 11 | > 帮助多线程的合理停止和异常处理 12 | - 实例 13 | ```python 14 | import tensorflow as tf 15 | import threading 16 | import time 17 | 18 | # 创建一个方法 通过coord来控制线程 19 | def MyLoop(coord): 20 | count=0 21 | while not coord.should_stop(): 22 | print("count=",count) 23 | count=count+1 24 | time.sleep(1) 25 | if count==10: 26 | print("stop") 27 | coord.request_stop() 28 | 29 | # Main thread: create a coordinator. 30 | coord = tf.train.Coordinator() 31 | 32 | # Create 3 threads that run 'MyLoop()' 33 | threads = [threading.Thread(target=MyLoop, args=(coord,)) for i in range(3)] 34 | for t in threads: 35 | t.start() # start a thread 36 | coord.join(threads) # wait util thread stop 37 | ``` 38 | 39 | ### tf.train.QueueRunner 40 | > 队列的执行者 41 | > 创建一系列线程，在同一个队列中组织tensors的排队执行 42 | - 实例 43 | ```python 44 | #1 create Runner 45 | b =tf.constant(2,dtype=tf.float16,shape=[1]) 46 | que=tf.RandomShuffleQueue(3,1,dtypes=[tf.float16],shapes=())#shapes一定要写 否则会报错 47 | enqueue_op =que.enqueue_many(b) 48 | 49 | qr = tf.train.QueueRunner(que,[enqueue_op]) 50 | #2 create Threads 51 | sess=tf.Session() 52 | coord = tf.train.Coordinator() 53 | enqueue_threads = qr.create_threads(sess, coord=coord, start=True) 54 | #3 use coord and queque 55 | inputs=que.dequeue_many(10)+b 56 | for step in range(10): 57 | if coord.should_stop(): 58 | break 59 | print("run") 60 | print(sess.run(inputs)) 61 | print("this request") 62 | coord.request_stop() 63 | coord.join(enqueue_threads)# wait thread to terminate 64 | ``` 65 | ![流程图](https://github.com/TFStudents/Tensorflow/blob/master/Resource/thread_queque.png) 66 | 67 | -------------------------------------------------------------------------------- /BaseLearn/threads/threadExample.py: -------------------------------------------------------------------------------- 1 | # python/user 2 | import tensorflow as tf 3 | import threading 4 | import time 5 | 6 | #---------------queque sample use------------------------ 7 | que=tf.FIFOQueue(3,[tf.int32]) 8 | node1= tf.constant([1,1]) 9 | que.enqueue_many(node1) 10 | #que.enqueue_many([122])# use the will make error 11 | print(que.dequeue()) 12 | print(que.dequeue()) 13 | print(que.dequeue()) 14 | 15 | 16 | #-----------------Coordinator sample use-------------------- 17 | def MyLoop(coord): 18 | count=0 19 | while not coord.should_stop(): 20 | print("count=",count) 21 | count=count+1 22 | time.sleep(1) 23 | if count==10: 24 | print("stop") 25 | coord.request_stop() 26 | 27 | # Main thread: create a coordinator. 28 | coord = tf.train.Coordinator() 29 | 30 | # Create 10 threads that run 'MyLoop()' 31 | threads = [threading.Thread(target=MyLoop, args=(coord,)) for i in range(3)] 32 | for t in threads: 33 | t.start() # start a thread 34 | coord.join(threads) # wait util thread stop 35 | 36 | #-----------------QueueRunner example------------------------------------ 37 | #1 create Runner 38 | b =tf.constant(2,dtype=tf.float16,shape=[1]) 39 | que=tf.RandomShuffleQueue(3,1,dtypes=[tf.float16],shapes=()) 40 | enqueue_op =que.enqueue_many(b) 41 | 42 | qr = tf.train.QueueRunner(que,[enqueue_op]) 43 | #2 create Threads 44 | sess=tf.Session() 45 | coord = tf.train.Coordinator() 46 | enqueue_threads = qr.create_threads(sess, coord=coord, start=True) 47 | #3 use coord and queque 48 | inputs=que.dequeue_many(10)+b 49 | for step in range(10): 50 | if coord.should_stop(): 51 | break 52 | print("run") 53 | print(sess.run(inputs)) 54 | print("this request") 55 | coord.request_stop() 56 | coord.join(enqueue_threads)# wait thread to terminate 57 | 58 | 59 | 60 | -------------------------------------------------------------------------------- /BaseLearn/variable/README.md: -------------------------------------------------------------------------------- 1 | # 变量 2 | [源文件](./variable.py) 3 | 地址[tf.Variable](https://www.tensorflow.org/api_docs/python/tf/Variable) 4 | ```Python 5 | #文件头先引入 6 | import tensorflow as tf 7 | ``` 8 | 9 | ## 如何创建？ 10 | - 构造函数 11 | ```Python 12 | __init__( 13 | initial_value=None, 14 | trainable=True, 15 | collections=None, 16 | validate_shape=True, 17 | caching_device=None, 18 | name=None, 19 | variable_def=None, 20 | dtype=None, 21 | expected_shape=None, 22 | import_scope=None 23 | ) 24 | ``` 25 | 从构造函数中可以看出，可以不传入任何参数，但是必须传入initial_value的值 不然会报错 26 | 其他值可以任意组合传入 如下 27 | ```Python 28 | v1=tf.Variable(0) 29 | v2=tf.Variable("hello world") 30 | ``` 31 | - 机器配置 32 | ```Python 33 | # 将变量设置到cpu 0 34 | with tf.device("/cpu:0"): 35 | v = tf.Variable(...) 36 | 37 | #将变量设置到gpu 0 38 | with tf.device("/gpu:0"): 39 | v = tf.Variable(...) 40 | ``` 41 | 42 | ## 初始化 43 | 创建变量之后，需调用tf.global_variables_initializer()进行初始化 44 | 才能进行调用 不然会报错的 45 | ```Python 46 | #初始化方式1 47 | init=tf.global_variables_initializer()#变量的初始化 初始化当前module所有的变量 48 | sess=tf.Session() 49 | sess.run(init) 50 | print(v1)#Tensor("Variable/read:0", shape=(), dtype=int32) 51 | #初始化方式2 52 | v2=tf.Variable(2) 53 | print(sess.run(v2.initialized_value()))#由于global_variables_initializer 其实不写也可以的 54 | print(v2)# Tensor("Variable_1/read:0", shape=(), dtype=int32) 55 | ``` 56 | ## 保存及读取变量 57 | ```Python 58 | # 保存session 59 | saver=tf.train.Saver() 60 | saver.save(sess,"./save") 61 | v1=v1.assign(12) 62 | print(sess.run(v1))#重新赋值后打印12 63 | 64 | a=saver.restore(sess,"./save")# 读取模型 65 | print(sess.run("v1:0"))#读取已经保存的变量v1 打印为0 66 | # saver=tf.train.import_meta_graph("save.meta") 67 | # saver.restore(sess,tf.train.latest_checkpoint("./")) 68 | # print(sess.run("v1:0"))# 可以从内存中读取变量v1 打印为0 69 | ``` 70 | 71 | -------------------------------------------------------------------------------- /BaseLearn/variable/variable.py: -------------------------------------------------------------------------------- 1 | #关于variable 2 | import tensorflow as tf 3 | 4 | v1=tf.Variable(0,name="v1")#变量的创建 5 | 6 | # 将变量设置到cpu 0 7 | #with tf.device("/cpu:0"): 8 | # v = tf.Variable(...) 9 | #将变量设置到gpu 0 10 | #with tf.device("/gpu:0"): 11 | # v = tf.Variable(...) 12 | #初始化方式1 13 | init=tf.global_variables_initializer()#变量的初始化 初始化当前module所有的变量 14 | sess=tf.Session() 15 | sess.run(init) 16 | print(v1)#Tensor("Variable/read:0", shape=(), dtype=int32) 17 | #初始化方式2 18 | v2=tf.Variable(2) 19 | print(sess.run(v2.initialized_value()))#由于global_variables_initializer 其实不写也可以的 20 | print(v2)# Tensor("Variable_1/read:0", shape=(), dtype=int32) 21 | # 保存session 22 | saver=tf.train.Saver() 23 | saver.save(sess,"./save") 24 | v1=v1.assign(12) 25 | print(sess.run(v1))#重新赋值后打印12 26 | 27 | a=saver.restore(sess,"./save")# 读取模型 28 | print(sess.run("v1:0"))#读取已经保存的变量v1 打印为0 29 | # saver=tf.train.import_meta_graph("save.meta") 30 | # saver.restore(sess,tf.train.latest_checkpoint("./")) 31 | # print(sess.run("v1:0"))# 可以从内存中读取变量v1 打印为0 -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2017 TFStudents 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /PicClassify/Newer/README.md: -------------------------------------------------------------------------------- 1 | # 直接运行程序 2 | 我们先不用看概念，不必知道里面的原理，直接将程序运行起来，看看效果，感受一下手写字体的识别。 3 | 1. ### 训练数据的下载 4 | > [地址](http://yann.lecun.com/exdb/mnist/) 下载如图绿色框内的4个数据，请确保是源文件，不要修改后缀名称 5 | ![图片](https://github.com/TFStudents/TF/blob/master/Resource/%E8%AE%AD%E7%BB%83%E9%9B%86.png) 6 | 在本地创建一个文件夹如MNIST_TEST（最好和代码同级目录）,将下载的4个文件放在其中。 7 | 2. ### 运行py代码 8 | > [mnist_softmax.py源码下载](...) 9 | > 将此源码下载保存到刚才新建的MNIST_TEST 同级目录下。在命令行，直接运行命令 10 | ``` 11 | python mnist_softmax.py 12 | ``` 13 | 在ipyhon控制台会有如下类似log 14 | ``` 15 | runfile('C:/WorkspaceOther/mnist_softmax.py', wdir='C:/WorkspaceOther') 16 | Extracting MNIST_TEST\train-images-idx3-ubyte.gz 17 | Extracting MNIST_TEST\train-labels-idx1-ubyte.gz 18 | Extracting MNIST_TEST\t10k-images-idx3-ubyte.gz 19 | Extracting MNIST_TEST\t10k-labels-idx1-ubyte.gz 20 | 0.9192 21 | ``` 22 | 使用ipython命令行，可能有如下错误，这是由ipython自身问题，可以忽略。 23 | ``` 24 | An exception has occurred, use %tb to see the full traceback. 25 | 26 | SystemExit 27 | 28 | C:\DevelopTools\Anaconda\envs\tensorflow\lib\site-packages\IPython\core\interactiveshell.py:2870: UserWarning: To exit: use 'exit', 'quit', or Ctrl-D. 29 | warn("To exit: use 'exit', 'quit', or Ctrl-D.", stacklevel=1) 30 | ``` 31 | # 解释代码及原理 32 | > 直接根据就运行顺序开始走起，引包就不讲了 33 | - ### 训练集文件地址 34 | ```python 35 | FLAGS = None 36 | DEFAULT_PATH="MNIST_TEST" 37 | ``` 38 | 由于源文件和训练集是同级目录，所以只需“MNIST_TEST”，如果不在，则这里填写训练集路径就可以了 39 | 40 | - ### 读取参数并启动main 41 | ``` 42 | if __name__ == '__main__': 43 | parser = argparse.ArgumentParser() #就是对命令行的参数进行解析及添加 44 | parser.add_argument('--data_dir', type=str, default=DEFAULT_PATH, 45 | help='Directory for storing input data') 46 | FLAGS, unparsed = parser.parse_known_args() 47 | tf.app.run(main=main, argv=[sys.argv[0]] + unparsed) #启动main(_)函数 48 | ``` 49 | [argparse.ArgumentParser更多知识参考](https://docs.python.org/2/howto/argparse.html) 50 | 51 | - ### 读取数据 52 | ``` 53 | mnist = input_data.read_data_sets(FLAGS.data_dir, one_hot=True) 54 | ``` 55 | 这才是真正的读取数据，tf已经封装了，其实FLAGS.data_dir的值为训练集的文件夹即"MNIST_TEST"，如果本地为空，系统会自动帮你下载，感兴趣可以看看源码。 56 | ### 至于数据集本身，是由一定像素的如28x28的手写数字图片，转换而成的特殊数据类型，至于怎么制作这样的图片，制作属于自己的数据集，这是个问题，再另外的章节再一起探讨。我有这样的计划，再每学一个实例的时候，先把基本的流程及核心原理弄懂，然后对应疑惑的问题进行整理，后面再深入的研究探讨。并且在章节的末尾会留下问题，如果你对某个问题感兴趣，可以提交出来，让大家学习参考。* 57 | 58 | - ### 创建模型 59 | ``` 60 | x = tf.placeholder(tf.float32, [None, 784])#None代表可以是任何长度 61 | W = tf.Variable(tf.zeros([784, 10])) 62 | b = tf.Variable(tf.zeros([10])) 63 | y = tf.matmul(x, W) + b #线性方程 64 | 65 | y_ = tf.placeholder(tf.float32, [None, 10])# 占位符 66 | ``` 67 | 你可以理解为平面中的直线方程,只是这里是多维的。机器学习的目标就是找到这一个一组参数来描述我们的数字特征。 68 | 为什么这里是784？是因为我们的输入28x28=784的像素。 69 | 为什么是10维？是因为数字0-9总共10个。 70 | x与W相乘会得到一个 与b一样的长度的一维向量。【None代表可以是任何长度】 71 | 至于这个方程代表什么含义，y代表的输出0-9,而y的右边代表输入经过神经网络的矩阵计算。 72 | 如何更深地理解此方程，请参考地址： 73 | [感知机](https://hit-scir.gitbooks.io/neural-networks-and-deep-learning-zh_cn/content/chap1/c1s1.html)和 74 | [sigmoid神经元](https://hit-scir.gitbooks.io/neural-networks-and-deep-learning-zh_cn/content/chap1/c1s2.html) 75 | 看完后你可能就会对这个方程有很深的认识。 76 | 77 | - ### 错误的评估之交叉熵 78 | ``` 79 | cross_entropy = tf.reduce_mean( 80 | tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y)) 81 | train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy) 82 | ``` 83 | y_是数据集中已经确定的数，y是评估模型，就是目标函数，是一个表达式，cross_entropy是交叉熵，用来评估错误率，这个值越小，代表错误率越低. 84 | -------------------------------------------------------------------------------- /PicClassify/Newer/mnist_softmax.py: -------------------------------------------------------------------------------- 1 | # Copyright 2015 The TensorFlow Authors. All Rights Reserved. 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | # ============================================================================== 15 | 16 | """A very simple MNIST classifier. 17 | 18 | See extensive documentation at 19 | https://www.tensorflow.org/get_started/mnist/beginners 20 | """ 21 | from __future__ import absolute_import 22 | from __future__ import division 23 | from __future__ import print_function 24 | 25 | import argparse 26 | import sys 27 | 28 | from tensorflow.examples.tutorials.mnist import input_data 29 | 30 | import tensorflow as tf 31 | 32 | FLAGS = None 33 | DEFAULT_PATH="MNIST_TEST" 34 | 35 | def main(_): 36 | # Import data 37 | mnist = input_data.read_data_sets(FLAGS.data_dir, one_hot=True) 38 | 39 | # Create the model 40 | x = tf.placeholder(tf.float32, [None, 784]) 41 | W = tf.Variable(tf.zeros([784, 10])) 42 | b = tf.Variable(tf.zeros([10])) 43 | y = tf.matmul(x, W) + b # 44 | 45 | # Define loss and optimizer 46 | y_ = tf.placeholder(tf.float32, [None, 10]) 47 | 48 | # The raw formulation of cross-entropy, 49 | # 50 | # tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(tf.nn.softmax(y)), 51 | # reduction_indices=[1])) 52 | # 53 | # can be numerically unstable. 54 | # 55 | # So here we use tf.nn.softmax_cross_entropy_with_logits on the raw 56 | # outputs of 'y', and then average across the batch. 57 | cross_entropy = tf.reduce_mean( 58 | tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y)) 59 | train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy) 60 | 61 | sess = tf.InteractiveSession() 62 | tf.global_variables_initializer().run() 63 | # Train 64 | for _ in range(1000): 65 | batch_xs, batch_ys = mnist.train.next_batch(100) 66 | sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys}) 67 | 68 | # Test trained model 69 | correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1)) 70 | accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) 71 | print(sess.run(accuracy, feed_dict={x: mnist.test.images, 72 | y_: mnist.test.labels})) 73 | writer = tf.summary.FileWriter("test_1", sess.graph) 74 | 75 | if __name__ == '__main__': 76 | parser = argparse.ArgumentParser() 77 | parser.add_argument('--data_dir', type=str, default=DEFAULT_PATH, 78 | help='Directory for storing input data') 79 | FLAGS, unparsed = parser.parse_known_args() 80 | tf.app.run(main=main, argv=[sys.argv[0]] + unparsed) -------------------------------------------------------------------------------- /PicClassify/README.md: -------------------------------------------------------------------------------- 1 | # 图片识别 2 | 3 | 图片识别用于识别图片属于哪一类物体,本质是图片的分类问题(机器学习主要用于两方面,回归和分类),此处以MNIST为例以数字识别介绍图片识别. 4 | 5 | ## 1. 环境 6 | python 2.7 7 | tensorflow 1.2.1 8 | 9 | ## 2. 程序说明 10 | 程序使用 tensorflow branch r1.2 中的https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/examples/learn/mnist.py 11 | 12 | $ ./minst.py 13 | 本程序采用两种方式进行训练,一种为LinearClassifier,另一种为cnn,其中cnn有进行两层卷积,下面进行详细说明. 14 | 15 | ### 2.1 LinearClassifier 16 | ```python 17 | feature_columns = learn.infer_real_valued_columns_from_input( 18 | mnist.train.images) 19 | classifier = learn.LinearClassifier( 20 | feature_columns=feature_columns, n_classes=10) 21 | classifier.fit(mnist.train.images, 22 | mnist.train.labels.astype(np.int32), 23 | batch_size=100, 24 | steps=1000) 25 | score = metrics.accuracy_score(mnist.test.labels, 26 | list(classifier.predict(mnist.test.images))) 27 | print('Accuracy: {0:f}'.format(score)) 28 | ``` 29 | LinearClassifier较易理解,不进一步说明 30 | 31 | ### 2.2 cnn 32 | 33 | cnn 部分主要包括卷积层,池化层和全连接层,详细介绍如下: 34 | 35 | #### 2.2.1 池化层 36 | ```python 37 | def max_pool_2x2(tensor_in): 38 | return tf.nn.max_pool( 39 | tensor_in, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') 40 | ``` 41 | 42 | 池化使用过滤器为大小为2*2, 使用max_pooling,长和宽的步长均为2,padding方式为'SAME',所以池化后图片大小变为原来的一半. 43 | 44 | #### 2.2.2 卷积层 45 | ```python 46 | def conv_model(feature, target, mode): 47 | """2-layer convolution model.""" 48 | # Convert the target to a one-hot tensor of shape (batch_size, 10) and 49 | # with a on-value of 1 for each one-hot vector of length 10. 50 | target = tf.one_hot(tf.cast(target, tf.int32), 10, 1, 0) 51 | 52 | # Reshape feature to 4d tensor with 2nd and 3rd dimensions being 53 | # image width and height final dimension being the number of color channels. 54 | feature = tf.reshape(feature, [-1, 28, 28, 1]) 55 | 56 | # First conv layer will compute 32 features for each 5x5 patch 57 | with tf.variable_scope('conv_layer1'): 58 | h_conv1 = layers.convolution2d( 59 | feature, 32, kernel_size=[5, 5], activation_fn=tf.nn.relu) 60 | h_pool1 = max_pool_2x2(h_conv1) 61 | 62 | # Second conv layer will compute 64 features for each 5x5 patch. 63 | with tf.variable_scope('conv_layer2'): 64 | h_conv2 = layers.convolution2d( 65 | h_pool1, 64, kernel_size=[5, 5], activation_fn=tf.nn.relu) 66 | h_pool2 = max_pool_2x2(h_conv2) 67 | # reshape tensor into a batch of vectors 68 | h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64]) 69 | 70 | # Densely connected layer with 1024 neurons. 71 | h_fc1 = layers.dropout( 72 | layers.fully_connected( 73 | h_pool2_flat, 1024, activation_fn=tf.nn.relu), 74 | keep_prob=0.5, 75 | is_training=mode == tf.contrib.learn.ModeKeys.TRAIN) 76 | 77 | # Compute logits (1 per class) and compute loss. 78 | logits = layers.fully_connected(h_fc1, 10, activation_fn=None) 79 | loss = tf.losses.softmax_cross_entropy(target, logits) 80 | 81 | # Create a tensor for training op. 82 | train_op = layers.optimize_loss( 83 | loss, 84 | tf.contrib.framework.get_global_step(), 85 | optimizer='SGD', 86 | learning_rate=0.001) 87 | 88 | return tf.argmax(logits, 1), loss, train_op 89 | 90 | ``` 91 | 该卷积层使用两层卷积,卷基层使用的过滤器大小为5*5,padding方式为'SAME',激活函数使用relu,损失函数为SGD; 92 | * 第一层卷积输入为28*28,卷积后的大小为28*28,设置输出的深度为32;然后经过池化层处理后深度为32,大小为14*14; 93 | * 第二层卷积输入为14*14,卷积后的大小为14*14,深度为64,经过池化处理后深度仍为64,大小为7*7 94 | * 在两层卷积和池化后加一个全连接层,全连接层使用1024核. 95 | * 在全连接后使用softmax进行分类 96 | 97 | #### 2.2.3 cnn调用 98 | ```python 99 | ### Convolutional network 100 | classifier = learn.Estimator(model_fn=conv_model) 101 | classifier.fit(mnist.train.images, 102 | mnist.train.labels, 103 | batch_size=100, 104 | steps=20000) 105 | score = metrics.accuracy_score(mnist.test.labels, 106 | list(classifier.predict(mnist.test.images))) 107 | print('Accuracy: {0:f}'.format(score)) 108 | ``` 109 | 110 | ### 2.3 运行结果 111 | LinearClassifier方法在测试集上正确率为0.921600,cnn在测试集上正确率为0.967500 112 | 113 | 114 | __以上为个人理解,如有错误请指正,感谢!!!__ 115 | -------------------------------------------------------------------------------- /PicClassify/__init__.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/python 2 | # encoding: utf-8 -------------------------------------------------------------------------------- /PicClassify/mnist.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/python 2 | # encoding: utf-8 3 | 4 | """This showcases how simple it is to build image classification networks. 5 | It follows description from this TensorFlow tutorial: 6 | https://www.tensorflow.org/versions/master/tutorials/mnist/pros/index.html#deep-mnist-for-experts 7 | """ 8 | 9 | from __future__ import absolute_import 10 | from __future__ import division 11 | from __future__ import print_function 12 | 13 | import numpy as np 14 | from sklearn import metrics 15 | import tensorflow as tf 16 | 17 | layers = tf.contrib.layers 18 | learn = tf.contrib.learn 19 | 20 | 21 | def max_pool_2x2(tensor_in): 22 | return tf.nn.max_pool( 23 | tensor_in, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') 24 | 25 | 26 | def conv_model(feature, target, mode): 27 | """2-layer convolution model.""" 28 | # Convert the target to a one-hot tensor of shape (batch_size, 10) and 29 | # with a on-value of 1 for each one-hot vector of length 10. 30 | target = tf.one_hot(tf.cast(target, tf.int32), 10, 1, 0) 31 | 32 | # Reshape feature to 4d tensor with 2nd and 3rd dimensions being 33 | # image width and height final dimension being the number of color channels. 34 | feature = tf.reshape(feature, [-1, 28, 28, 1]) 35 | 36 | # First conv layer will compute 32 features for each 5x5 patch 37 | with tf.variable_scope('conv_layer1'): 38 | h_conv1 = layers.convolution2d( 39 | feature, 32, kernel_size=[5, 5], activation_fn=tf.nn.relu) 40 | h_pool1 = max_pool_2x2(h_conv1) 41 | 42 | # Second conv layer will compute 64 features for each 5x5 patch. 43 | with tf.variable_scope('conv_layer2'): 44 | h_conv2 = layers.convolution2d( 45 | h_pool1, 64, kernel_size=[5, 5], activation_fn=tf.nn.relu) 46 | h_pool2 = max_pool_2x2(h_conv2) 47 | # reshape tensor into a batch of vectors 48 | h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64]) 49 | 50 | # Densely connected layer with 1024 neurons. 51 | h_fc1 = layers.dropout( 52 | layers.fully_connected( 53 | h_pool2_flat, 1024, activation_fn=tf.nn.relu), 54 | keep_prob=0.5, 55 | is_training=mode == tf.contrib.learn.ModeKeys.TRAIN) 56 | 57 | # Compute logits (1 per class) and compute loss. 58 | logits = layers.fully_connected(h_fc1, 10, activation_fn=None) 59 | loss = tf.losses.softmax_cross_entropy(target, logits) 60 | 61 | # Create a tensor for training op. 62 | train_op = layers.optimize_loss( 63 | loss, 64 | tf.contrib.framework.get_global_step(), 65 | optimizer='SGD', 66 | learning_rate=0.001) 67 | 68 | return tf.argmax(logits, 1), loss, train_op 69 | 70 | 71 | def main(unused_args): 72 | ### Download and load MNIST dataset. 73 | mnist = learn.datasets.load_dataset('mnist') 74 | 75 | ### Linear classifier. 76 | feature_columns = learn.infer_real_valued_columns_from_input( 77 | mnist.train.images) 78 | classifier = learn.LinearClassifier( 79 | feature_columns=feature_columns, n_classes=10) 80 | classifier.fit(mnist.train.images, 81 | mnist.train.labels.astype(np.int32), 82 | batch_size=100, 83 | steps=1000) 84 | score = metrics.accuracy_score(mnist.test.labels, 85 | list(classifier.predict(mnist.test.images))) 86 | print('Accuracy: {0:f}'.format(score)) 87 | 88 | ### Convolutional network 89 | classifier = learn.Estimator(model_fn=conv_model) 90 | classifier.fit(mnist.train.images, 91 | mnist.train.labels, 92 | batch_size=100, 93 | steps=20000) 94 | score = metrics.accuracy_score(mnist.test.labels, 95 | list(classifier.predict(mnist.test.images))) 96 | print('Accuracy: {0:f}'.format(score)) 97 | 98 | 99 | if __name__ == '__main__': 100 | tf.app.run() -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Tensorflow 实例 2 | ## 基础 3 | 4 | - [安装](BaseLearn/install) 5 | - [简单上手](BaseLearn/start)  6 | - 变量 7 | > [创建-初始化](BaseLearn/variable)  8 | > [共享变量](BaseLearn/share_variable) 9 | - [多线程和队列](BaseLearn/threads) 10 | - [读取数据](BaseLearn/read_data) 11 | - [训练助手](BaseLearn/Supervisor) 12 | - TensorBoard的使用 13 | > 1. [显示graph](BaseLearn/tensorboard) 14 | > 2. [显示图片](BaseLearn/tensorboard_image) 15 | - 调试 16 | > [Mnist为例](https://www.tensorflow.org/programmers_guide/debugger)-[未完]   17 | > [机器学习为例](https://www.tensorflow.org/programmers_guide/tfdbg-tflearn) 18 | - [元图](https://www.tensorflow.org/programmers_guide/meta_graph) 19 | - [保存模型](https://www.tensorflow.org/programmers_guide/saved_model_cli) 20 | - [版本解释](https://www.tensorflow.org/programmers_guide/version_semantics) 21 | - [数据版本](https://www.tensorflow.org/programmers_guide/data_versions) 22 | 23 | ## 初级实例 24 | [参考资料](https://www.tensorflow.org/get_started) 25 | > ### 跑实例时 一步步解析每一行代码，包括每一个变量，每一个函数 每一步的运行结果写在后面注释里面 26 | - [mnist新手](./PicClassify/Newer/) 27 | - [mnist专家](./PicClassify) 28 | - [神经网络](https://www.tensorflow.org/get_started/mnist/mechanics) 29 | - 机器学习 30 | > [快速入门](https://www.tensorflow.org/get_started/tflearn) 31 | > [输入函数](https://www.tensorflow.org/get_started/input_fn) 32 | > [日志和监测](https://www.tensorflow.org/get_started/monitors) 33 | - TensorBoard 34 | > [可视化学习](https://www.tensorflow.org/get_started/summaries_and_tensorboard) 35 | > [嵌入可视化](https://www.tensorflow.org/get_started/embedding_viz) 36 | > [图像化](https://www.tensorflow.org/get_started/graph_viz) 37 | > [直方图](https://www.tensorflow.org/get_started/tensorboard_histograms) 38 | ## 算法相关 39 | - [GPU的使用](https://www.tensorflow.org/tutorials/using_gpu)[聂彬-认领] 40 | - [图片识别](./PicClassify)[张夏旭-认领]-完成   41 | - [迁移学习](https://www.tensorflow.org/tutorials/image_retraining)-[魏杨-认领] 42 | - 卷积神经网络-[魏杨-认领] 43 | >  [创建](https://www.tensorflow.org/tutorials/layers) 44 | > [高级](https://www.tensorflow.org/tutorials/deep_cnn) 45 | - [词的向量化](https://www.tensorflow.org/tutorials/word2vec)[聂彬-认领] 46 | - [循环神经网络](https://www.tensorflow.org/tutorials/recurrent)[聂彬-认领] 47 | - [序列到序列模型](https://www.tensorflow.org/tutorials/seq2seq)[张夏旭-认领] 48 | - [大型线性模型](https://www.tensorflow.org/tutorials/linear) 49 | - [线性模型实例](https://www.tensorflow.org/tutorials/wide) 50 | - [深度学习](https://www.tensorflow.org/tutorials/wide_and_deep)-[王凯-认领] 51 | - [Mandelbrot Set](https://www.tensorflow.org/tutorials/mandelbrot)-[王凯-认领]   52 | - [偏微分方程](https://www.tensorflow.org/tutorials/pdes)-[王凯-认领] 53 | 54 | 55 | [参考资料](https://www.tensorflow.org/programmers_guide) 56 | ### 本项目为开源项目：由[Tensorflow学生会]管理更新，TF爱好者参与 57 | - 一个DNN的分类例子 58 | 介绍内容：构造一个四层的DNN，用Python读取一份分类数据，并feed给DNN。一步步解析每一行代码，包括每一个变量，每一个函数 59 | - 一个CNN的分类例子（MNIST） 60 | 介绍内容：给出代码，并一步步解析每一行代码，包括每一个变量，每一个函数。 61 | - 转换image成tfrecord的例子 62 | - 读tfrecords，并训练网络使用tfrecords 63 | 64 | 65 | 66 | 67 | 68 | 69 | 70 | 71 | 72 | 73 | - end 其他需要修改的 大家提出来 74 | 75 | -------------------------------------------------------------------------------- /Resource/README.md: -------------------------------------------------------------------------------- 1 | 这是图片资源 2 | -------------------------------------------------------------------------------- /Resource/thread_queque.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/Resource/thread_queque.png -------------------------------------------------------------------------------- /Resource/训练集.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TFStudents/TensorFlowExample/6278f8b3d2c61e45c95ec47a10fa47e61472bf55/Resource/训练集.png --------------------------------------------------------------------------------