在windows10环境安装配置tensorflow1.2及初步学习

此处有详细说明：https://github.com/tensorflow/tensorflow/tree/v1.2.0/tensorflow/contrib/cmake

1.安装anacoda3，即安装python即相关库：https://www.continuum.io/downloads#windows（python为3.6版本）

2.从git中下载tensorflow源码：https://github.com/tensorflow/tensorfl：ow

3.下载windows版本的swig：http://www.swig.org/download.html，本例解压到C盘

4.进入tensorflow源码目录：D:\tensorflow-1.2.0\tensorflow\contrib\cmake，新建build文件夹，进入build后，并执行命令如下（本例为CPU版）

cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release -DSWIG_EXECUTABLE=C:/swigwin-3.0.12/swig.exe -DPYTHON_EXECUTABLE=C:/ProgramData/Anaconda3/python.exe -DPYTHON_LIBRARIES=C:/ProgramData/Anaconda3/libs/python36.lib

cmake中就会卡住，原因不明

注：

原来按照tensorflow官方的windows方法安装，使用anacoda下的pip进行安装，但无法针对CPU的SSE等指令集进行优化（没有GPU情况下），故采用源码方法安装。

但据说https://stackoverflow.com/questions/42603407/how-to-compile-tensor-flow-with-sse-and-and-avx-instructions-on-windows

即使cpu优化 也没有什么好处， 故放弃这种方法

Tensorflow学习笔记

20170710

TensorFlow’shigh-level machine learning API (tf.contrib.learn)

训练及识别过程：

1.   Load CSVs containing Iris training/testdata into a TensorFlowDataset

2.   Construct aneural network classifier

3.   Fit the model using the training data

4.   Evaluate the accuracy of the model

5.   Classify new samples

Where isos.path?

 

# 打开数据集

load_csv_with_header()method in learn.datasets.base，返回值为dataset

数据集格式：Csv文件

第一行特殊：样本的个数特征向量的维度类别的名称

余下若干行：特征向量的内容类别标签

dataset是named tuples，training_set.data与training_set.target分别表示特征向量和标签。

# 定义网络结构

#Specify that all features have real-value data
feature_columns =[tf.contrib.layers.real_valued_column("", dimension=4)]
# Build 3 layer DNN with 10, 20, 10 unitsrespectively.
classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
                                          hidden_units=[10,20,10],
                                          n_classes=3,
                                          model_dir="/tmp/iris_model")

# 开始训练

#Define the training inputs
def get_train_inputs():
  x = tf.constant(training_set.data)
  y = tf.constant(training_set.target)
  return x, y

# Fitmodel.
classifier.fit(input_fn=get_train_inputs, steps=2000)

#验证结果

#Define the test inputs
def get_test_inputs():
  x = tf.constant(test_set.data)
  y = tf.constant(test_set.target)
  return x, y

# Evaluate accuracy.
accuracy_score =classifier.evaluate(input_fn=get_test_inputs,
                                    steps=1)["accuracy"]

print("\nTest Accuracy: {0:f}\n".format(accuracy_score))

 

针对数据集训练迭代4000次loss为0.0287，测试集准确率为0.966

 

※Custom Input Pipelines withinput_fn

defmy_input_fn():
    # Preprocessyour data here...
    # ...thenreturn 1) a mapping of feature columns to Tensors with
    # thecorresponding feature data, and 2) aTensor containing labels
    returnfeature_cols, labels

输出前，应该将数组转变为Tensor[tf.constant(…)]

或者SparseTensor[tf.SparseTensor(…)](三参数定义dense_shape, indices, values)

 

※在尝试cwTH训练时侯出现错误

raise type(e)(node_def, op, message)

更改DNNClassifier中model_dir目录即可，否则会在这目录训练好的基础上继续训练

训练结果显示四层（56，48，32，16）隐层的网络两万次迭代可以实现训练集90+的结果，但泛化能力差

 

20170711

C API 中包括了 run apredefined graph的功能：

A language binding isexpected to define the following classes:

·Graph: A graph representing a TensorFlow computation. Consists ofoperations (represented in the client language by Operations) and corresponds to a TF_Graph in the C API. Mainly used as an argument when creatingnew Operation objects and when starting a Session. Also supports iterating through the operations in the graph (TF_GraphNextOperation), looking up operations by name (TF_GraphOperationByName), and converting to and from a GraphDef protocol message (TF_GraphToGraphDef and TF_GraphImportGraphDef in the C API).

·Operation: Represents a computation node in the graph. Corresponds toa TF_Operation in the C API.

·Output: Represents one of the outputs of an operation in the graph.Has a DataType (and eventually a shape). May be passed as an inputargument to a function for adding operations to a graph, or to a Session's Run()method to fetchthat output as a tensor. Corresponds to a TF_Output in the C API.

·Session: Represents a client to a particular instance of the TensorFlowruntime. Its main job is to be constructed with a Graph and some options and then field calls to Run() the graph. Corresponds to a TF_Session in the C API.

·Tensor: Represents an N-dimensional (rectangular) array with elementsall the same DataType.Gets data into and out of a Session's Run() call.Corresponds to a TF_Tensor inthe C API.

·DataType: An enumerant with all the possible tensor types supported byTensorFlow. Corresponds to TF_DataType in the C API and often referred to as dtype in the Python API.