If we could take a low res sample of the network's hidden layers, with the right sampling algorithm it might be possible to compare them and see how they relate to each other.
If so, it might tell us all kinds of useful info about the labels used in the training data, let us visualise them, group, graph, find interesting intersections, areas of the network that
| // | |
| // Uncrops Google Docs images by removing the clip path. | |
| // Doesn't work in Webkit due to canvas elements not being | |
| // in the dom. So Firefox only for now. | |
| // | |
| // Click the "raw" button and import the URL into your userscript | |
| // manager - tested in tampermonkey but should work in whatever. | |
| // | |
| // ==UserScript== |
| #!/usr/bin/python | |
| import json | |
| import logging | |
| import os | |
| import re | |
| import urllib | |
| import urllib2 | |
| from subprocess import call |
| git clone [email protected]:mozilla/gecko-dev.git | |
| cd gecko-dev | |
| for year in $(seq 2007 2024); do | |
| echo $year | |
| git log --since="$year-01-01" --until="$year-12-31" --pretty=format:"%ae" | tr '[:upper:]' '[:lower:]' > $year.txt | |
| done | |
| blacklist="grep -Ev [email protected]|[email protected]|automation\-bot" |