After wasting many hours picking images for albums and artists in my streaming music application, bemused, I started writing an application to help me pick the images. It took only a few seconds to realize, I have a the makings of a project that could benefit from machine learning. I have now built up a training set of images I have selected and if I can generate a set of images I reject, then I have built up a perfect training set for an algorithm that could make recommendations or even decide. I've avoided applying ML to things that don't really need it, but this is a classic problem.
This project will have a few phases that will be tracked in this repository:
For all my various datasets, I have these steps to do. I will start with the Music site.
- Phase 1: Generate our training dataset
- Phase 2: Discover and test some image algorithms
- Phase 3: Use the algorithm to pick real images for existing Bemused data
- Phase 4: Determine how good we did
- I would like to automate the way I make decisions about choosing an image for an artist.
- I have about 1,000 artists with images available. I have 9,000 without an image.
- I need to add images I rejected to my training set
- I would like to auto-favorite tracks.
- I have a few years of logs of music that's been streamed on my website.
- I need to augment my data with some more features about the music itself
- I would like to auto-generate playlists. Based on my favorite music
- I will definitely need more information about my music.
- I would like to deduplicate artists in my library. (may be a non-ML task)
- I would like to use get suggestions for what I could listen to.
- I would like to automate recommendations on posts I should read from weblogs, Hacker News, Lobsters
- I have years of saved websites from my personal caching site, pigeon
- I have a bunch of favorites at Hacker News
- I would like to prioritize my inbox and find messages I missed over the years
- I have years of emails I have read and responded to for training
- I would like to automate the way I choose images for my music.
- I would like to deduplicate artists in my library. (may be an ML task)
Bemused (formerly PShare) is a personal music project I've been working on since 2003. In the beginning, I created it as a simple way for me to download my mp3 files on new computers. Over time, I bought a flash plugin that allowed me to stream music to my devices. This clearly better. As html5 developed, I rewrote the site in Ruby and used JPlayer to use the streaming capabilities that became native parts of browser apis. Somewhere in the 2000's, I started adding support for images. At first, I acquired artwork from Amazon when I would import mp3 files. Later, I added supprt for manually overriding or setting these images. As of this writing, I have about a thousand artist images out of ten thousand in my system. There are a large number of duplicates, but I still have a huge amount of artists without images.