conceptual-machine

just early hacking so far

"we derive CAVs by training a linear classifier between a concept’s examples and random counter examples and then taking the vector orthogonal to the decision boundary"

"a CAV encodes the direction of a concept in the vectorspace of a bottleneck"

"compute cosine similarity between a set of picturesof interest to the CAV to sort the pictures"

screenshots

flow

debugging log for nerds

watch the training (thanks tfvis!)

these animals are cute but not very shiba...

(activations from concept classifier projected with umap, colored by concept classification)

these ones are moar shiba

huh, the scottish terrier concept seems weird...

how does the model do on pictures of terriers?

let's look over in warping-machine:

woah guess it is pretty biased

...

datasets

oxford pets

https://www.robots.ox.ac.uk/~vgg/data/pets/ https://www.kaggle.com/tanlikesmath/the-oxfordiiit-pet-dataset

random scripts

glasses/not, from celeba

# find offset (add 1 to it)
cat annotations/list_attr_celeba.txt | tail -n +2 | head -n1 | cut -d' ' -f16

# 17: glasses
rm -rf sampled/eyeglasses-no
rm -rf sampled/eyeglasses-yes
rm -rf sampled/random-tail

mkdir -p sampled/eyeglasses-no
mkdir -p sampled/eyeglasses-yes
mkdir -p sampled/random-tail

# sample 40 each (from head)
cat annotations/list_attr_celeba.txt | tail -n +3 |  awk '{ print $1, $17 }' | grep ' 1' | cut -d' ' -f1 | sort -R | head -n40 | xargs -I {} cp img_align_celeba/{} sampled/eyeglasses-yes/{}
cat annotations/list_attr_celeba.txt | tail -n +3 |  awk '{ print $1, $17 }' | grep ' -1' | cut -d' ' -f1 | sort -R | head -n40 | xargs -I {} cp img_align_celeba/{} sampled/eyeglasses-no/{}

# random sample (from tail)
cat annotations/list_attr_celeba.txt | tail -n +3 |  awk '{ print $1 }' | sort -R | tail -r | head -n40 | xargs -I {} cp img_align_celeba/{} sampled/random-tail/{}