With the advent of tensorflow 2.0, can pytoch still shake the status of big brother?

TensorFlow 2.0 preview

About TensorFlow 2.0 preview, Open source strategist at Google Edd Wilder-James An email that was to be made public introduced ,TensorFlow 2.0 The preview will be officially released this year , And called it a major milestone . Will focus on ease of use , and Eager Execution Will be TensorFlow 2.0 Core functions .

notes ：“Eager Execution” It's an imperative 、 Interface defined by run , Once from Python Called to perform the operation immediately , This makes
TensorFlow It's easier to get started with , It also makes R & D more intuitive .

TensorFlow 2.0 go online

TensorFlow 2.0 preview Finally online , It seems that the stable version is not too far away from us —— Officially, it will be the first quarter of the year .
Google said , In the last few years ,TensorFlow Added a lot of components . adopt TensorFlow 2.0 A massive reconstruction of the version , These functions will be packaged into an integrated platform , Support the entire machine learning workflow from training to deployment . The figure below shows briefly TensorFlow 2.0 The new architecture of ：

Note： Although the training section above focuses on Python API, however TensorFlow.js It also supports training models .TensorFlow 2.0 There is also varying degrees of support for other languages , Include Swift、R Language and Julia.

Market share

The global situation

2.0 The press conference also put TensorFlow My current background has been revealed ： at present TF There have been more than 4100 Million downloads , The community has more than 1800 Multiple contributors .

A global map will be displayed at the press conference , But there is no disclosure about the Chinese community , How can this be ？

Official picture

An overview of China

Here is my Chinese search engine – Baidu index statistics through the search data to observe the two mainstream deep learning framework tensorflow And pytorch The change of , The results are as follows ：

The picture above shows the two mainstream frameworks of deep learning in China in the past year tensorflow And pytorch Contrast between , It is obvious that
tensorflow Far better than pytorch. Especially in the presence of TF2.0 After the announcement , The search index gap has widened .

from Crowd attributes Come up and say 20~29 as well as 30-39 Between the crowd , Younger people are more likely to pytorch, Older programmers are more inclined to tensorflow.

TensorFlow And PyTorch difference

1. Installation environment
   First of all, support on the system ：all in . But it's worth noting that 2018Pytorch v0.4.0 Support windows Platform .

2. CPU and GPU
   TensorFlow Targeted CPU and GPU Install the module , and PyTorch Don't like TensorFlow The same has been specified CPU and GPU, If you want to support GPU and CPU, More code will be generated .

3. setup script
   be based on Anaconda Both of the two deep learning modules can be directly passed through Pip To install .

4. Whether it's suitable for beginners
   TensorFlow 1.x And PyTorch By contrast , Personally think that PyTorch many , But in tensorflow 2.0 After release, according to its new features ,Tensorflow 2.0 Will be in PyTorch Be roughly the same .

The following is a comparison of some specific aspects ：
PyTorch And TensorFlow 1
For example, to calculate 1 + ½ + ¼ + ⅛ + … , Use PyTorch The code is obviously better than TensorFlow Simple ：

Later from TensorFlow 1.4 Start , You can choose to start eager Pattern .

stay TensorFlow 2.0, eager execution By default , It doesn't need to be enabled ：

You can find eager Patterns and PyTorch It's as simple as .

5. In effect
   I think it's for different needs 、 Different algorithms have different choices , There is no absolute good or bad .

TensorFlow2.0 New characteristics

Let's take a look at 2.0 New features of version ：2.0 The version is simple 、 Clearer 、 Three characteristics of expansibility , Greatly simplify API; Improved TensorFlow Lite and TensorFlow.js The ability to deploy models ;

TensorFlow2.0 Alpha To sum up, i.e ：

• Easier to use ：

Such as tf.keras Wait for advanced API Will be easier to use ; also Eager execution Will be the default setting .

• Clearer ：

Removed duplicate features ; Different API Call syntax is more consistent 、 intuitive ; Better compatibility .

• More flexible ：

Provide complete low level API; Can be found in tf.raw_ops Access internal operations ; Provide variables 、checkpoint Inheritable interfaces to and layers .

A brief summary of the main changes

• API clear
  many API stay TF 2.0 To disappear or move . Some of the major changes include the deletion of tf.app,tf.flags And tf.logging, Open source support absl-py（Google Their own Python The code base ）.

• Eager Execution Will become the core function

  Probably TensorFlow 2.0 The most obvious change is to make Eager execution As the default priority mode . This means that any operation will run immediately after the call , We no longer need to predefine static diagrams , Re pass 「tf.Session.run()」 All parts of the executive chart .

    # TensorFlow 1.X
    outputs = session.run(f(placeholder), feed_dict={placeholder: input})
    # TensorFlow 2.0
    outputs = f(input)

• Code style with Keras Mainly

Many functions such as optimizer,loss,metrics Will be integrated into Keras in

• Support more platforms and languages

1.0 To 2.0 transition

Automatic transition

About code conversion ： from TensorFlow1.0 To 2.0 Transition we use pip install TensorFlow 2.0 when , The system will automatically add tf_upgrade_v2（ project Address ） , It can take the existing TensorFlow Python Code conversion to TensorFlow 2.0 Code .

# Usage method ：
!tf_upgrade_v2

#  choice  input file, Output  output file
tf_upgrade_v2 --infile foo.py --outfile foo-upgraded.py

#  Transform the entire directory 
tf_upgrade_v2 --intree coolcode --outtree coolcode-upgraded

Compatibility

To ensure TensorFlow 2.0 Still support your code , The upgrade script contains a compat.v1 modular . This module replaces TF
1.x Symbol tf.foo, Equivalent to tf.compat.v1.foo The reference is the same . Although the compatibility module is very good , But we recommend that you manually proofread the replacement and migrate it to tf. New in the namespace API, instead of tf.compat.v1..

because TensorFlow
2.x Module is obsolete （ for example ,tf.flags and tf.contrib）, So switch to compat.v1 Some can't fix changes . Upgrading this code may require the use of other libraries （ for example absl.flags） Or switch to tensorflow
/ addons In the package .

   The above is from the official website 

Summary

TensorFlow 2.0 A very powerful and mature deep learning library has been simplified , The point is to keras Mainly , I wonder if you understand keras, According to the official slogan , It is “ Design for human beings , It's not designed for machines API”. So he will be greatly optimized in terms of entry , If you have the following needs , that TensorFlow Is a good choice ：

• Develop models that need to be deployed on mobile platforms
• Want rich learning resources in various forms （TensorFlow There are many development courses ）
• Want or need to use Tensorboard
• Large scale distributed model training is needed

PyTorch Still a young framework , But it's growing faster and faster . If you have the following needs , It might be better for you ：

• Rapid prototyping for small scale projects
• To study

reference

https://github.com/tensorflow/docs/blob/master/site/en/r2/guide/effective_tf2.md
https://tensorflow.google.cn/
https://www.youtube.com/watch?v=WTNH0tcscqo&t=304s