<AI Startup> Programming Challenge

Author: Yona Edell

A random RGB image generator that creates & displays images using multiprocessing, numpy, and opencv. Created as part of a timed programming challenge given only these instructions.

Getting Started

• Written with Python 3.8.8
  • will probably work with Python 3.6+ (but don't hold me to that 😛)
• Install dependencies: pip install -r requirements.txt

Run main program:

python colors.py

Run tests:

pytest test_colors.py

Usage

When running colors.py, you should see the following prompt on the command line:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<     Random Image Creator & Viewer     >
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Enter number of images to generate:

User is prompted for input to number of images to generate, image width, and image height. Once the user submits the input, a new window opens to display the randomly generated RGB images. The user can press the Enter key to view the next image, or the Q key to quit the program.

Bonus feature: text size & position watermarked on the image are relative(ish) to image dimensions/size!

Important Notes:

• Display/text/app scaling on your OS will affect this program's image viewing dimensions
  • turn off if you want to view large images at their true resolution; instructions for Windows 10 here
• If you're viewing the last image and you press Enter (to view next image), the program will auto-quit
  • if this behavior is undesired, i.e. you'd like to be forced to press Q to quit after the last image, comment out Line 282 in colors.py
• The size of your RAM is the only real limit to how many images you can generate ✨
  • that's right: I generated 100,000 500x500px images and let the program use 8GB+ of my system's memory before pressing Q--it took ~30 seconds to flush the queues & completely shutdown, but it got there!

Conclusion

I learned a ton throughout this project, especially since it was my first time using multiprocessing, opencv, or pytest (I've only used the unittest library). I hit several snags along the way and spent too much time diagnosing race conditions & deadlocks and manually killing zombie/orphaned processes. The most challenging parts were keeping array_a synchronized to view/get the images in order from queue_b and making a clean exit when quitting before all queue items had been consumed. Honorable mentions go to the horse's ass behind the opencv BGR color space choice (this one got me good) and perpetually mixing up the width/height dimensions between the numpy, multiprocessing, and opencv arrays.

As far as results go, I think the testing aspect has the most room for improvement. I'm not satisfied with how much my simple/functional unit tests cover-- I left out a unit test for the display_image() function since it's inherently convoluted (the best I could do would be checking that it terminates, but how much good is that?). I'd really like to refactor/clean up most of my tests (use fixtures, setup/teardowns, classes), but I think I'll go to bed now! 😴

To-do

• Refactor code to be more DRY, especially the tests
• Eliminate one of the multiprocessing Events; probably don't need event_start, could maybe use a Condition
• Use process pooling/queue managers or joinable queues so don't have to worry about queue flushing upon quit
• Add more test cases and different kinds of tests (performance, stress, end-to-end/integration, etc)
• Add more timeouts to shared objects (to be more defensive in preventing deadlocks)
• Add more/fancier colors!!!
• Add throttling or I/O caching functionality so memory doesn't get hogged when generating thousands of images
• Use threads instead of processes to reduce resource consumption & potentially add speedup (?)
• Make "auto-quit" feature (mentioned above in Important Notes) a command line option with argparse