/MATH_32_HW2

Primary LanguagePython

MATH_32_HW2

NOTE

For all questions related to this homework, please open a github issue so others with similar problems can benefit from it.

Or you can contact the instructors via email: Shen Zheng, Changjie Lu

Overview

In this assignment, you will implement a basic style transfer framework utilizing Gram matrices. The primary objectives are understanding how style and content can be manipulated within images and gaining experience in hyperparameter tuning. This task will enhance your comprehension of style transfer principles and aid in the development of your optimization skills.

Installation

  • Check HW1 for basic installations.
  • Install Pillow using pip install Pillow
  • Install Git and connect it to VS Code. Here is a tutorial.

Code Completion

  • Open style_and_content.py, and complete 5 TODO's.
  • The 1st and 2nd TODOs ask you to implement the content loss.
  • The 3rd TODO asks you to implement the gram matrix.
  • The 4th and 5th TODOs ask you to implement the style loss using the gram matrix.

Hyperparameter Tuning

  • Open run.sh and uncomment the lines under ############# Hyperparameter Tuning ##############
  • Run in terminal 
    bash run.sh

Exchange Content and Style

  • Open run.sh and uncomment the lines under ############# Exchange Content and Style ##############
  • Run in terminal 
    bash run.sh

Style Transfer (Given Images)

  • Open run.sh and uncomment the lines under ############# Style Transfer ##############
  • Run in terminal 
    bash run.sh

Style Transfer (Your Images)

  • Prepare a content image and a style image on your own.
  • Open run.sh and add a line like this: 
    run_main $path_to_your_content_image $path_to_your_style_image
  • Run the following command in the terminal: 
    bash run.sh

Results and Discussion

  • For Hyperparameter Tuning
    • Which style weight (1e4, 1e5, 1e6) produces the best visual result?
    • Why isn't a larger style weight always preferable for style transfer?
  • For Exchange Content and Style
    • How does the image appear when you exchange content and style?
    • Can you explain the types of images suitable for content/style?
  • For Style Transfer
    • Does the style transfer result look visually pleasing?
    • If not, list at least three ways to improve it.

Submission

  • NOTE: Updated 6/20/2023
  • Navigate to the submission folder for HW2 at here
  • Locate the folder with your name.
  • Include all the results in a PDF file.
  • The PDF file should contain the following information, from top to bottom:
    • Your name, email, and student ID.
    • Codes for the five TODOs.
    • Images for ALL experiments (each row: content image, style image, and output image)
    • Short answers for ALL questions in the Results and Discussion section.
  • Submit the PDF file by placing it in your folder.

Deadline

  • 07/02/2023 07/07/2023