1
$\begingroup$

What is a conditioning vector?

Context: from the StackGAN paper about generating realistic pictures from text descriptions

$\endgroup$
1
$\begingroup$

The paper probably gives the best explanation. The idea is that rather than using the CNN-RNN embeddings directly as inputs to the generator, you want to reprocess them to a smaller size (the raw embeddings are length 1024 but the conditioning vectors are length 128) and also add some randomness/sample more diverse examples from embedding space. You do this by learning 128 means and stddevs from the raw embeddings using a FC layer, then essentially sampling the 128 components of the conditioning vector from 128 independent normal distributions with those means and stddevs

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.