CS 20SI|Lecture 1 Introduction to TensorFlow

本文是基于课程CS 20SI: TensorFlow for Deep Learning Research进行总结。

课程目标

• 理解TF计算图方法
• 探索TF内建函数
• 学习如何构建适合DL项目的结构化模型

参考书籍

• TensorFlow for Machine Intelligence (TFFMI)
• Hands-On Machine Learning with Scikit-Learn and TensorFlow. Chapter 9: Up and running with TensorFlow
• Fundamentals of Deep Learning. Chapter 3: Implementing Neural Networks in TensorFlow (FODL)
  TensorFlow正在快速迭代，上述书籍可能会过时，直接参考官方网站

开始

High level API

1. TF Learn
2. TF Slim
3. Keras(官方宣布支持作为上层API)
4. Tensor Layer

Graphs and Sessions

Data Flow Graphs

TensorFlow将计算图的定义从执行过程中分离。TF的工作分为两个阶段：
阶段1：定义计算图
阶段2：使用一个session执行图中的operations

Tensor

Tensor是一个n维矩阵，相当于numpy中的ndarray

import tensorflow as tfa = tf.add(3, 5)

上述代码构建了上面的数据流图，在数据流图中结点Node可以是operators, variables, and constants,边Edge是tensors
Tensor Flow即为Tensor(Data)在图中流动。

import tensorflow as tfa = tf.add(3, 5)print(a)>> Tensor("Add:0", shape=(), dtype=int32)

注意到print(a)并没有输出结果5.下面介绍如何得到a的数值

Session

Session对象封装了Operation对象执行和Tensor对象evaluated的环境

import tensorflow as tfa = tf.add(3, 5)sess = tf.Session()print(sess.run(a))sess.close()>> 8

Session会查看计算图中的结点，寻找如何能够得到a的结果，并计算沿途的nodes.
使用上下文管理器可以不用每次显式关闭Session.

import tensorflow as tfa = tf.add(3, 5)with tf.Session() as sess:    print(sess.run(a))

More (sub) graph

tf.Session.run(fetches, feed_dict=None, options=None, run_metadata=None)
将想要计算的变量作为一个list传递给fetches

import tensorflow as tfx = 2y = 3op1 = tf.add(x, y)op2 = tf.mul(x, y)useless = tf.mul(x,op1)op3 = tf.pow(op2, op1)with tf.Session() as sess:    op3, not_useless = sess.run([op3,useless])

Distributed Computation

将计算图拆分适合于多机多核并行计算

Another Graph

• Multiple graphs require multiple sessions, each will try to use all available resources by default
• Can’t pass data between them without passing them through python/numpy, which doesn’t work in distributed
• It’s better to have disconnected subgraphs within one grap

g = tf.Graph()#to add operators to a graph, set it as default:with g.as_default():    x = tf.add(3, 5)sess = tf.Session(graph=g)with tf.Session() as sess:    sess.run(x)

小心不要混淆默认graph和用户定义graph,向任何计算图添加操作时，先将其定义为默认图

g1 = tf.get_default_graph()g2 = tf.Graph()#add ops to the default graphwith g1.as_default():    a = tf.Constant(3)#add ops to the user created graphwith g2.as_default():    b = tf.Constant(5)

Why graphs

1. Save computation (only run subgraphs that lead to the values you want to fetch)
2. Break computation into small, differential pieces to facilitates auto-differentiation
3. Facilitate distributed computation, spread the work across multiple CPUs, GPUs, or devices
4. Many common machine learning models are commonly taught and visualized as directed graphs already