Advantages Of using Pytorch | Coursera Community
Coursera Header
Solved

Advantages Of using Pytorch

  • 28 July 2019
  • 3 replies
  • 210 views

Userlevel 3
Badge +5
Pytorch has become immensely popular. I'm seeing lot of people moving from Tensorflow/Keras to Pytorch and many articles recommending it. Whenever I search for machine learning models implemented from scratch, most of the results will be in pytorch. Why is pytorch's popularity suddenly increasing? What are the advantages of using pytorch? For those who are using pytorch, what benefits did you found in pytorch which were not there in tensorflow/Keras?
icon

Best answer by Mo Rebaie 29 July 2019, 17:07

Hello @THANGA MANICKAM M, it's true that a lot of people are moving from Tensorflow to PyTorch, but also there are a lot of people moving from PyTorch to Tensorflow, so it's not a rule to follow. Moreover, Tensorflow is also becoming immensely popular and many articles recommending it.

I have learned to work with both PyTorch and Tensorflow, the 2 frameworks are powerful to implement "complex" DL projects and both are created by AI pioneers, so there is no "certain" advantage for any of these 2 frameworks compared with the other one.

It depends on the learner, if you see yourself more flexible to learn Tensorflow then work with it, and same for PyTorch.
View original

3 replies

Userlevel 6
Badge +5
Hello @THANGA MANICKAM M, it's true that a lot of people are moving from Tensorflow to PyTorch, but also there are a lot of people moving from PyTorch to Tensorflow, so it's not a rule to follow. Moreover, Tensorflow is also becoming immensely popular and many articles recommending it.

I have learned to work with both PyTorch and Tensorflow, the 2 frameworks are powerful to implement "complex" DL projects and both are created by AI pioneers, so there is no "certain" advantage for any of these 2 frameworks compared with the other one.

It depends on the learner, if you see yourself more flexible to learn Tensorflow then work with it, and same for PyTorch.
Userlevel 3
Badge +5
Thanks for sharing @Mo Rebaie
Userlevel 2
Badge +1
In my view, the drastic advantage of Pytorch was the eager evaluation model, that makes debugging so much easier (as opposed to horrible). But since rel. 2.0, TF has eager execution by default. And meanwhile Pytorch got serving and graph mode execution (with Torchscript). The choice is harder than ever. Perhaps check on LinkedIn, what is the most required for jobs you could be interested in?

Then if I Google for "Pytorch eager execution" most of the top results relate to Tensorflow instead; it is an obviously rigged result, which makes me want to adopt Pytorch just to spite Google.

Reply