Good method for parallelizing in python?

Question

I have a quite simple task. I have a video and first i have to slice it per frame, predict it using tensorflow, and combine those predicted frames into a single video. Here's the result I got:

Slicing ... Time elapsed since last slice : 0s. Frame-1
.
Slicing ... Time elapsed since last slice : 2.0517446994781494s. Frame-2
.
Slicing ... Time elapsed since last slice : 0.8912339210510254s. Frame-3
.
Slicing ... Time elapsed since last slice : 0.8657193183898926s. Frame-4
.
Slicing ... Time elapsed since last slice : 0.8655312061309814s. Frame-5
.
Slicing ... Time elapsed since last slice : 0.8827650547027588s. Frame-6
.
Slicing ... Time elapsed since last slice : 0.8690683841705322s. Frame-7
.
Slicing ... Time elapsed since last slice : 0.8906550407409668s. Frame-8
.
Slicing ... Time elapsed since last slice : 0.8798754215240479s. Frame-9
.
Slicing ... Time elapsed since last slice : 0.9341959953308105s. Frame-10

each frame took about 0.8 second. Let's say i have a 5 seconds video with 30 fps. It will take about 120 seconds to process.

I have a thought about parallelizing for this like multithreading etc. Is this even possible? Where do I have to start? Thank you.

Not easily. Python cant truly multithread because it has a Global Interpreter Lock (GIL). It can multi-task/multi-thread but that doesn't speed anything up because the threads can't run at the same time they just keep alternating turns. Here is a good article about the GIL Grok the GIL: How to write fast and thread-safe Python — Error - Syntactical Remorse
– Error - Syntactical Remorse, Commented Sep 26, 2019 at 16:16
The article above also goes over what is needed for Parallelism and how it works. — Error - Syntactical Remorse
– Error - Syntactical Remorse, Commented Sep 26, 2019 at 16:22
@Error-SyntacticalRemorse That's not exactly true – multiple threads can run native code (or anything that releases the GIL). Only Python code is run in a quasi-cooperative-multitasking sort of manner. — AKX
– AKX, Commented Sep 26, 2019 at 16:31
@AKX I tried to emphasize that with the "Not easily". It is possible but not easy to multi thread to multiple cores. It is possible to truly multi-thread (such as NumPy) but not easily. Any native code that uses the GIL though will not use multiple cores unless you make it so each thread gets a GIL which the link emphasizes. — Error - Syntactical Remorse
– Error - Syntactical Remorse, Commented Sep 26, 2019 at 16:33
For OP's usecase, they can easily use multiprocessing to spread out frame processing to multiple processes instead. — AKX
– AKX, Commented Sep 27, 2019 at 6:43

Community · Accepted Answer · 2020-06-20 09:12:55Z

1

You can try out multiprocessing package in python. For more details refer: https://docs.python.org/3/library/multiprocessing.html.

Or other thing you can try out is to Maximize TensorFlow Performance on CPU by optimization. The primary goal of the performance optimization is to make use of all the cores available in the machine and thereby speed up the process. This can be achieved by setting the intra_op_parallelism_threads, inter_op_parallelism_threads, OMP_NUM_THREADS etc.

Please refer : https://software.intel.com/en-us/articles/maximize-tensorflow-performance-on-cpu-considerations-and-recommendations-for-inference

Hope this helps.

edited Jun 20, 2020 at 9:12

CommunityBot

11 silver badge

answered Sep 26, 2019 at 16:28

Aswathy - Intel

6484 silver badges12 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Good method for parallelizing in python?

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related