This repository has been archived by the owner on Dec 3, 2020. It is now read-only.

92% price accuracy#275

Merged

Osmose merged 42 commits intomozilla:masterfrom erikrose:90%-price

Nov 20, 2018

+208-173

Commits on Nov 16, 2018

Get trainer running on image out-rule from webext-commerce.
erikrose
committed
Add a trainee for each out() rule so we can choose them from the Trainer menu.
erikrose
committed
Respell a regex for clarity.
erikrose
committed
Bring up to date with 9813ba8 in https://github.com/mozilla/price-wise .
erikrose
committed
Rewrite isAboveTheFold using trapezoid() and fuzzy-logic scores.
erikrose
committed
Refactor largerImage as well. This completes the image coefficients. Images still score 100%.
erikrose
committed
Rewrite image y-axis scorer to constrain to 0..1 and for simplicity.
erikrose
committed
Retrain to fix the priceish coeff for isAboutTheFold().
erikrose
committed
Change rules that look for "price" in IDs and classes to emit fuzzy confidences. Also the $ rule.
erikrose
committed
Re-express font-size rule as a confidence.
erikrose
committed
Rewrite rule that give a bonus to prices near the winning image.
erikrose
committed
Express hasPriceishPattern as a fuzzy truth.
erikrose
committed
Fix the bugs that immediately kept the trainer from training.
erikrose
committed
Remove a now-unused constant and an out-of-date comment.
erikrose
committed
Add new coeffs to get to 100% on the training set!
erikrose
committed
Rename hasPriceIn, since it doesn't actually have "price" hard-coded in anymore.
erikrose
committed
Consider divs with background images as well as img tags.
erikrose
committed
Typos
erikrose
committed
There's no need to say "node". All scoring functions take nodes.
erikrose
committed
Add a rule to punish extreme aspect ratios and another to punish background images. This gets us to 100% on the training corpus of 75.
erikrose
committed
Hard-code a height for aboveTheFold so a user's different window size doesn't make a different answer come out.
erikrose
committed
Make the image trainee train only the image-affecting coeffs, for speed.
erikrose
committed
Move tuned image coeffs into master vector.
erikrose
committed
Make a more efficient training vector for price.
erikrose
committed
Improve price coeffs: 100% on 12-16.
erikrose
committed
Improve price coeffs. 93.3% on 12-16, 1-10. 92% on 1-25.
erikrose
committed
Improve price coeffs: 98.7% on all current training samples (1-25 and all named samples).
erikrose
committed
Copy tuned price coeffs to master vector.
erikrose
committed
Put the glue code back how I found it, and move the coeffs back into the JSON file.
erikrose
committed
Fix a mistranscribed coefficient.
erikrose
committed
Make linter happy.
erikrose
committed

Commits

Commits on Nov 16, 2018

Commits on Nov 20, 2018