What is the difference between pytorch and tensorflow?

PyTorch Is a fully functional framework for building deep learning models , meanwhile tensorflow It is also one of the commonly used frameworks . When we are studying , Try it for comparison . that pytorch and tensorflow What's the difference? ？ Issues of concern , Here comes the answer .

pytorch and tensorflow What's the difference? ？

Creating and running a graph is probably the most different between the two frameworks . stay PyTorch in , The graph structure is dynamic , This means that diagrams are built at run time . And in the TensorFlow in , The graph structure is static , This means that the graph is first “ compile ” And then it runs again .

Let's take a simple example , stay PyTorch You can use the standard Python Write a grammar for Loop structure

for _ in range(T):
    h = torch.matmul(W, h) + b

here T Can be changed every time the code is executed . and TensorFlow in , This requires the use of “ Control flow operation ” To build a graph , for example tf.while_loop.TensorFlow It does provide dynamic_rnn For common structures , But creating custom dynamic calculations is really more difficult .

PyTorch The simple graph structure in is easier to understand , what's more , And easier to debug . debugging PyTorch Code is like debugging Python code . You can use pdb And set breakpoints anywhere . debugging TensorFlow Code is not easy . Or you have to request the variables to be checked from the session , Or learn to use TensorFlow The debugger (tfdbg).

flexibility

tensorflow： Static calculation diagram , The data parameters are in CPU And GPU Migration between is troublesome , Debugging trouble

pytorch： Dynamic calculation diagram , The data parameters are in CPU And GPU Migration between is very flexible , Easy debugging

Calculation speed

Under the same conditions ：

tensorflow stay CPU Upper running speed ratio pytorch fast

tensorflow stay GPU Running speed and pytorch almost

Dependency Library

tensorflow： Support more library functions , For example, there will be more image data preprocessing methods

pytorch： Expanding , The future and tensorflow It shouldn't be that different

Data loading

tensorflow：API Huge design , Use skillfully , But sometimes you can't load data directly into TensorFlow

pytorch：API The overall design is rough , But loading data API The design is very friendly . The interface for loading data consists of a data set 、 A sampler and a data loader constitute .

Equipment management

tensorflow： There's no need to manually adjust , Simple

pytorch： It is necessary to specify the enabled devices

On this point , By personal preference , I prefer to enable it manually CUDA_VISIBLE_DEVICES, My personal memory 、 The management of video memory is strict .

  Share some of my artificial intelligence learning materials for free , Including some AI Common framework actual combat video 、 Image recognition 、OpenCV、NLQ、 machine learning 、pytorch、 Computer vision 、 Videos such as deep learning and neural network 、 Courseware source code 、 Famous essence resources at home and abroad 、AI Hot papers 、 Industry reports, etc .

For better systematic learning AI, I recommend that you collect one .

Here are some screenshots , Free download method is attached at the end of the article .

One 、 Artificial intelligence courses and projects

Two 、 Famous essence resources at home and abroad

3、 ... and 、 Collection of papers on artificial intelligence

Four 、 AI Industry Report

Learn Artificial Intelligence well , Read more , Do more , practice , If you want to improve your level , We must learn to settle down and learn slowly and systematically , Only in the end can we gain something .

Click on the business card below , Scan the code and download the information for free .