一 TensorFlow GPU环境
环境和版本信息: 操作系统:ubuntu14.04
GPU:NIVIAD GTX 1070
Python版本:3.5以上
CUDNN:7.0
CUDA:9.0
TensorFlow-GPU:1.10.1 注:大家不一定要严格按照以上的版本,主要是保证CUDA、CUDNN与显卡能够匹配即可。
二TensorFlow三个核心概念
2.1Graph:
在TensorFlow中，计算图时一个有向图，用来描述计算节点以及计算节点之间的关系，所以在TensorFlow中我们存储在一个值活着数组的时候，存的其实是这个值活着数组的计算图而不是其本身的数字。
#####关于计算图的操作
1，新建计算图: g=tf.Graph(),但是不同计算图赏的张量是不能共享的，这个是存在于变量
2，指定计算图的使用的deice: with g.device("/gpu:0")
3, 设置默认计算图: with g.as_default:
4, 在绘画中可以指定使用的计算图: with tf.Session(graph=g1):
#GPU版
root@c3c3d78304f9:~# cat gpu_example.py
#!/usrbin/env python3.5
import tensorflow as tf
g = tf.Graph()
with g.device("/gpu:0"):
d = tf.constant([10,9,8,7])
e = tf.constant([1,2,3,4])
f_1 = d+e
print(f_1.graph)
print(d.graph,e.graph)
sess=tf.Session()
print(sess.run(f_1))
root@c3c3d78304f9:~# python3.5 gpu_example.py
2018-11-11 12:51:27.575737: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-11-11 12:51:27.678706: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:898] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-11-11 12:51:27.680526: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1356] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.721
pciBusID: 0000:29:00.0
totalMemory: 10.91GiB freeMemory: 10.55GiB
2018-11-11 12:51:27.680544: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1435] Adding visible gpu devices: 0
2018-11-11 12:51:27.843339: I tensorflow/core/common_runtime/gpu/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-11-11 12:51:27.843381: I tensorflow/core/common_runtime/gpu/gpu_device.cc:929] 0
2018-11-11 12:51:27.843388: I tensorflow/core/common_runtime/gpu/gpu_device.cc:942] 0: N
2018-11-11 12:51:27.843594: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 10210 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:29:00.0, compute capability: 6.1)
[11 11 11 11]
CPU 版本
import tensorflow as tf
a = tf.constant([1,2,3,4],name='a')
b = tf.constant([0,1,2,3],name='b')
c=a+b
print(a.graph, b.graph)
print(c.graph)
sess=tf.Session()
print(sess.run(c))
运行结果
[1 3 5 7]
import tensorflow as tf
g1=tf.Graph() #定义一张图
with g1.as_default():
a = tf.constant([1,2,3], name='a')#用常量试试看
b = tf.get_variable('b',initializer=tf.constant_initializer()(shape=[1]))#用常量试试看
g2=tf.Graph()
with g2.as_default():
a = tf.constant([2,3],name='a')#用常量试试看
b = tf.get_variable('b',initializer=tf.constant_initializer()(shape=[3]))#用常量试试看
＃在session里面指定图
with tf.Session(graph=g2) as sess:
with g1.device("/cpu:0"):
tf.global_variables_initializer().run()
c=a+1
print("常量的情况下 = ",sess.run(c))
with tf.variable_scope("",reuse=True):
print("变量情况下 = ",sess.run(tf.get_variable("b")))
＃在session里面指定图
with tf.Session(graph=g2) as sess:
with g2.device("/gpu:0"):
tf.global_variables_initializer().run()
c=a+1
print("常量的情况下 = ", sess.run(c))
with tf.variable_scope("",reuse=True):
print("变量情况下 = ",sess.run(tf.get_variable("b")))
运行结果
root@c3c3d78304f9:~# python3.5 1.py
2018-11-11 13:28:40.767842: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-11-11 13:28:40.870770: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:898] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-11-11 13:28:40.871278: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1356] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.721
pciBusID: 0000:29:00.0
totalMemory: 10.91GiB freeMemory: 10.55GiB
2018-11-11 13:28:40.871295: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1435] Adding visible gpu devices: 0
2018-11-11 13:28:41.063014: I tensorflow/core/common_runtime/gpu/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-11-11 13:28:41.063058: I tensorflow/core/common_runtime/gpu/gpu_device.cc:929] 0
2018-11-11 13:28:41.063069: I tensorflow/core/common_runtime/gpu/gpu_device.cc:942] 0: N
2018-11-11 13:28:41.063274: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 10210 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:29:00.0, compute capability: 6.1)
常量的情况下= [3 4]
变量情况下= [0. 0. 0.]
2018-11-11 13:28:41.148431: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1435] Adding visible gpu devices: 0
2018-11-11 13:28:41.148473: I tensorflow/core/common_runtime/gpu/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-11-11 13:28:41.148481: I tensorflow/core/common_runtime/gpu/gpu_device.cc:929] 0
2018-11-11 13:28:41.148487: I tensorflow/core/common_runtime/gpu/gpu_device.cc:942] 0: N
2018-11-11 13:28:41.148575: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 10210 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:29:00.0, compute capability: 6.1)
常量的情况下 = [3 4]
变量情况下 = [0. 0. 0.]
2.2张量
张量(tensor)可以简单理解为多维数组。其中零阶张量表示标量(scalar),也就是一个数；一阶张量为向量(vector),也就是一维数组；
第n阶张量可以理解为一个n维数组。但是张量在TensorFlow中的实现病不是直接采用数组的形式，它只是对TensorFlow中运算结果的引用。在张量中病没有真正保存数字，它保存的是如何得到这些数字的计算过程。
2.3定义张量的三种方式
import tensorflow as tf
2.3.1 用的比较多
with tf.Session as less:
print(less.run(result))
2.3.2
sess = tf.Session()
with less.as_default():
print(result.eval())
2.3.3交互式，不需要指定会话，默认
sess=tf.InteractiveSession()#会自动成为默认会话
会话结束后，所有的变量都消失
root@c3c3d78304f9:~# cat tensorflow_example.py
#!/usr/bin/env python3.5
import tensorflow as tf
a = tf.constant(2,name='a')
b = tf.constant([0,1,2,3], name='b')
c=a*b
print(a,b)
print(c)
sess = tf.Session()
print(sess.run(c))
root@c3c3d78304f9:~# python3.5 tensorflow_example.py
Tensor("a:0", shape=(), dtype=int32) Tensor("b:0", shape=(4,), dtype=int32)
Tensor("mul:0", shape=(4,), dtype=int32)
2018-11-11 09:57:47.148221: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-11-11 09:57:47.253357: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:898] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-11-11 09:57:47.254083: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1356] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.721
pciBusID: 0000:29:00.0
totalMemory: 10.91GiB freeMemory: 10.55GiB
2018-11-11 09:57:47.254112: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1435] Adding visible gpu devices: 0
2018-11-11 09:57:47.423734: I tensorflow/core/common_runtime/gpu/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-11-11 09:57:47.423776: I tensorflow/core/common_runtime/gpu/gpu_device.cc:929] 0
2018-11-11 09:57:47.423785: I tensorflow/core/common_runtime/gpu/gpu_device.cc:942] 0: N
2018-11-11 09:57:47.423987: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 10210 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:29:00.0, compute capability: 6.1)
[0 2 4 6]
root@c3c3d78304f9:~#
2.3Session:
在TensorFlow中，计算图的计算过程都是在会话下进行的，同一个会话内的数据是可以共享的，会话结束计算的中间量酒会小时。
在TensorFlow需要指定会话。
三TensorFlow常用基础API
根据TensorFlow官网以及在日常的编程中的使用情况，梳理了一下需要掌握的TensorFlow基础API:
tf.Graph()
tf.Graph.device()
tf.Graph.as_default()
tf.Session()
tf.Session.run()
tf.Session.as_default()
tf.InteractiveSession()
tf.constant()
tf.Variable()
tf.get_variable()
tf.placeholder() ＃占位，行参
tf.train() ＃重点
tf.nn()#重点
tf.nn.conv2d
tf.nn.relu
tf.nn.max_pool
tf.nn.dropout
tf.nn.softmax
tf.train
tf.train.GradientDescentOptimizer
tf.train.Saver
四重点API详解
4.1 tf.nn
tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, name=None)
第一个参数input: 指需要做卷积的输入图像，它要求是一个Tensor,具有[batch, in_height,in_width,in_channels]这样的shape,具体含义是［训练时一个batch的图片数量，图片高度，图片宽度，图像通道数］,注意这时一个4维的Tensor,要求类型为float32和float64其中之一
第二个参数filter:相当于CNN中的卷积核，它要求是一个Tensor,具有[filter_height,filter_width, in_channels, out_channels]这样的shape,具体含义是[卷积核的高度，卷积核的宽度，图像通道数，卷积核个数]，
要求类型与参数input相同，有一个地方需要注意，第三维in_channels,就是参数input的第四维
第三个参数strides:卷积时在图像每一维的补偿，这时一个一维的向量，长度4
第四个参数padding:string类型的量，只能是"SAME","VALID"其中之一
第五个参数: use_cudnn_on_gpu:bool类型，是否使用cudnn加速，默认为true
4.2 tf.nn.relu
tf.nn.relu(features, name=None)
这个函数的作用是计算激活函数 relu,即max(features,0).即将矩阵中每行的非最大值置为0
4.3 tf.nn.max_pool池化层
第一个参数value:需要池化的输入，一般池化层接在卷积层后面，所以输入通常是feature map, 依然是[batch,height,width,channels]这样的shape
第二个参数ksize:池化窗口的大小，取一个四维向量，一半时[1,height, width,1],因为我们不想在batch和channels上做池化，所以这两个维度设为了1
第三个参数strides:和卷积类似，窗口在每一个维度上滑动的步长，一般也是[1,stride,stride, 1]
第四个参数padding:和卷积类似，可以取'VALID' 或者'SAME'
4.4 tf.nn.dropout
tf.nn.dropout(x,keep_prob,noise_shape=None,seed=None,name=None)
x :输入tensor..
keep_prob: float类型，每个元素被保留下来的概率
noise_shape:一个1维的int32张量，代表了随机产生"保留/丢弃"标志的shape
seed :整形变量，随机数种子
4.5 tf.nn.softmax
标签编程两个其实就是逻辑回归，一般多标签（多图像识别）
五tf.train
tf.train
梯度下降
tf.train.GradientDescentOptimizer
优化器
tf.train.Saver
保存我们的模型
六项目和代码设计思路
Flask_app :前端等
execute :定义好model后，执行器负责预测，训练，调用model来做预测
cnnModel :定义一个类，把公共方法和参数都放入里面
getConfig :所有参数都可以配置，所有key都在配置文件中配置，修改起来非常方便，修改一个值所有都改掉
数据集下载地址
255 wget http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
256 wget http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
257 wget http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
258 wget http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz
