search

tensorflow / tcav

Code for the TCAV ML interpretability project
Apache License 2.0
628 stars 144 forks source link

result not reproduced by taking gradient wrt logit and flipping sign #124

Open avani17101 opened 2 years ago

avani17101 commented 2 years ago

I had the same doubt as https://github.com/tensorflow/tcav/issues/83 https://github.com/tensorflow/tcav/issues/107 https://github.com/tensorflow/tcav/issues/41 I flipped the sign for tcav calc (< changed to >)and took gradient wrt logit(unnormalized output of network, self.pred). Still, I am not getting the same tcav scores as when grad wrt loss and tcav calc sign being <

Why is this happening? @BeenKim ?

avani17101 commented 2 years ago

So the output for importance of zebra in inception network when wrt loss and < sign image whereas when wrt output and > sign I get this image