Is it possible enable multithreading in deepdream function?

[itsTeknas]

I've noticed that the actual deepdream function that does the work actually runs on a single thread. So no matter how powerful multicore processor you got, it will only use a very small portion of it.
I am a newbie in this field, can someone tell if its possible to have a multithreaded function ?

I executed htop inside the compute container and i got this... notice only core 1 is used and others are idle.

  1  [|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||100.0%]     5  [                                                                     0.0%]
  2  [                                                                     0.0%]     6  [                                                                     0.0%]
  3  [                                                                     0.0%]     7  [                                                                     0.0%]
  4  [                                                                     0.0%]     8  [                                                                     0.0%]
  Mem[|||||||||||||||||||||||||||||||||||||||||||||||||             1140/2001MB]     Tasks: 7, 1 thr; 2 running
  Swp[|||||||||||||||||                                               103/460MB]     Load average: 1.17 1.06 0.87 
                                                                                     Uptime: 05:00:11

  PID USER      PRI  NI  VIRT   RES   SHR S CPU% MEM%   TIME+  Command
  528 root       20   0 1633M 1037M 42900 R 100. 51.8  0:42.06 python deepdream.py
  544 root       20   0 23020  3292  2632 R  0.7  0.2  0:00.01 htop
    1 root       20   0 17960     0     0 S  0.0  0.0  0:00.07 /bin/bash -c cd /opt/deepdream && ./process_images.sh 2>&1 > log.html
    6 root       20   0 17964     0     0 S  0.0  0.0  0:00.00 /bin/bash ./process_images.sh
    7 root       20   0 17980     0     0 S  0.0  0.0  0:00.00 /bin/bash ./process_images_once.sh
    9 root       20   0 17984  1916  1784 S  0.0  0.1  0:00.04 /bin/bash ./process_images_once.sh
  529 root       20   0 1633M 1037M 42900 S  0.0 51.8  0:00.00 python deepdream.py
  530 root       20   0 18168  3300  2844 S  0.0  0.2  0:00.09 bash





F1Help  F2Setup F3SearchF4FilterF5Tree  F6SortByF7Nice -F8Nice +F9Kill  F10Quit

[jontonsoup]

bump

The deepdream process is inherently serialized. You have a stack of layers that the image gets fed through. The image goes into layer one, gets processed, then moves on to layer two, gets processed, gets moved to layer three, etc. You can't work on layers 1, 2, 3, and 4 simultaneously, as 4 requires the output from 3, 3 requires the output from 2, and so on. You could work on 4 different images concurrently if you wanted, but I'm not sure that is what you were getting at.

If you did want to work on 4 images concurrently, given that you have 4 cores, you would want multiprocessing, not multithreading, as CPU bound threading in python is gimped due to the GIL.

[jontonsoup]

Thanks-- I implemented multiprocessing already, the problem is I run into memory constraints so it's not great. I read about an alternate caffe build that enabled multi cores, but I haven't explored it yet.

On Mon, Aug 17, 2015 at 6:03 PM, ihaventkilledanybodysince1984
notifications@github.com wrote:

The deepdream process is inherently serialized. You have a stack of layers that the image gets fed through. The image goes into layer one, gets processed, then moves on to layer two, gets processed, gets moved to layer three, etc. You can't work on layers 1, 2, 3, and 4 simultaneously, as 4 requires the output from 3, 3 requires the output from 2, and so on. You could work on 4 different images concurrently if you wanted, but I'm not sure that is what you were getting at.

If you did want to work on 4 images concurrently, given that you have 4 cores, you would want multiprocessing, not multithreading, as CPU bound threading in python is gimped due to the GIL.

Reply to this email directly or view it on GitHub:
#34 (comment)

[StevenGann]

"Too bad there isn't some API that could distribute workload across GPU processors and do the same thing on CPUs depending on what was available"

I'm no expert, but I believe both CUDA and OpenCL are capable of this. I believe OpenCL can even manage it by default.