ugotsoul/fsdl_project

an applied ML project for the full-stack deep learning class

Ideas

Two topics I am interested replicating from wandb benchmarks:

• super resolution: https://wandb.ai/wandb/superres/benchmark

• poverty mapping with satellite data: https://wandb.ai/thinkingmachines/ph-poverty-mapping/benchmark

Other topics I'm interested in:

• Building and deploying a GAN similar to using my own drawings or other open sourced art I've gathered as a dataset

  • Ideas for datasets: protest signs, propaganda posters, gritty (eg., japanese product or sports mascots?), african masks

• Object detection:

  • Curb cuts for accessiblity or piles of trash from satellite imagery
  • Using a webcam video source: poor posture, face touching, or cleanched jaw.
  • Using a webcam: hornet activity (or bird activity?) in backyard

• Handwriting recognition / OCR

  • Graffiti text
  • Husband's writing

• Super resolution

  • Old cartoons
  • Vintage clothing
  • Drawings

Alternate projects:

• Contribute to fastai library
• Contribute to stylegan / stylegan-ada fork

Tooling:

• Mlflow
• Docker
• AWS/GCP
• Streamlit

Proposal

Guidelines: https://docs.google.com/document/d/1wyDSJsunIlheSRXiGCVEnrqYU7RPyvAX2mkAwmMKlMk/edit
Example proposal: https://docs.google.com/document/d/1w5czuaR87WIxF6PXwsxnoguzRIIyrCZk0yY_com0QaE/edit

Project Name:

Team: Ugaso Sheikh-Abdi

Project Goal:

What dataset will you use, or how do you plan to collect data?

What baseline will you use?

What model architecture / loss function do you propose?

What will the end result look like?

What’s your stretch goal: