search

piotrhm / image_recognition_project

0 stars 0 forks source link

Decorrelation #27

Closed panpiort8 closed 2 years ago

panpiort8 commented 2 years ago

Uf

solves #18

DorotaNowak commented 2 years ago

Good progress! Do you see any difference in the optimized images between this and previous experiments?

mproszewska commented 2 years ago

I'm sceptical about not passing gradients via generator. Can you link something that you based your code on? quote or something

adriansuwala commented 2 years ago

it's adapted from lucid https://github.com/tensorflow/lucid/blob/6dcc927e4ff4e7ef4d9c54d27b0352849dadd1bb/lucid/optvis/param/spatial.py#L48-L87 https://github.com/tensorflow/lucid/blob/6dcc927e4ff4e7ef4d9c54d27b0352849dadd1bb/lucid/optvis/param/color.py#L32

mproszewska commented 2 years ago

it's adapted from lucid https://github.com/tensorflow/lucid/blob/6dcc927e4ff4e7ef4d9c54d27b0352849dadd1bb/lucid/optvis/param/spatial.py#L48-L87 https://github.com/tensorflow/lucid/blob/6dcc927e4ff4e7ef4d9c54d27b0352849dadd1bb/lucid/optvis/param/color.py#L32

I can't find lines with gradient computations here

panpiort8 commented 2 years ago

it's adapted from lucid https://github.com/tensorflow/lucid/blob/6dcc927e4ff4e7ef4d9c54d27b0352849dadd1bb/lucid/optvis/param/spatial.py#L48-L87 https://github.com/tensorflow/lucid/blob/6dcc927e4ff4e7ef4d9c54d27b0352849dadd1bb/lucid/optvis/param/color.py#L32

I can't find lines with gradient computations here

Me neither. And I'm also sceptical about how it's implemented.

piotrhm commented 2 years ago

👍

mproszewska commented 2 years ago

Why is this merged? @panpiort8 didn't provide any proof that this method is just a way of determining initial image (which seems unlikely)

piotrhm commented 2 years ago

This solutions seems to not be what we expected but @panpiort8 provided new initialisation technique which "do something". As I understand such method is also provided in lucid and it seems reasonable to explore this topic a bit more - in such context merged code is a good starting point. Imho we should at least brainstormed it during next planning. Worst case scenario we delete it.