-
Notifications
You must be signed in to change notification settings - Fork 19
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
A simple question about the code #8
Comments
@KID0031 hey thanks for your interest yes you are right that a linear projection would probably do fine there as well. i'm following the weight tied output embedding technique from earlier transformer architectures (which in theory should allow the network to learn better embeddings), but that has been shown to be unnecessary i'll make it an option to do it the way you describe |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi, @lucidrains, I'm a beginner trying to use Q-transformer and encountered a question while reading the code. In the
QHeadMultipleActions
class, I noticed that Q-transformer encodes the bin into an embedding usingself.action_bin_embeddings
. However, when obtaining the q value, it multiplies the attention output withself.action_bin_embeddings
once again. Is there a specific reason for using this approach to derive the q value instead of employing a new MLP layer multiplied by the attention output? I've shared the relevant code below. Thank you!The text was updated successfully, but these errors were encountered: