-
Notifications
You must be signed in to change notification settings - Fork 83
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
More details about the "gather embeddings"? #5
Comments
For example, use the following prompt and target tokens. For clarity, I've also added the token number to the prompt.
NB. The comma is an independent token. For this prompt and target tokens, Cutoff (this extension) generates the three prompts shown below.
CLIP converts these three prompts into three corresponding vectors (each a 77x768 matrix):
where For instance, From the above, we have a vector to be applied to each token.
The embedding to be applied is: Sorry for my long text. I used the word "gather" with the intention of the action of picking one specific vector for each |
Just still a little confused about the term Cutoff. In this case, you actually cut off the current token. However, from the overall discription, it seem to cut off the other words that may affect the current one. |
i print the log for Cutoff, the "gather embeddings" method is clear explained by the @hnmr293 , i think the cut off words should be the way in the log, as @asheroin said below is the log info a cute girl, white hair, _ shirt, _ tie, _ shoes, _ short pants, _ eyes a cute girl, _ hair, black shirt, _ tie, _ shoes, _ short pants, _ eyes a cute girl, _ hair, _ shirt, green tie, _ shoes, _ short pants, _ eyes a cute girl, _ hair, _ shirt, _ tie, red shoes, _ short pants, _ eyes a cute girl, _ hair, _ shirt, _ tie, _ shoes, blue short pants, _ eyes a cute girl, _ hair, _ shirt, _ tie, _ shoes, _ short pants, yellow eyes |
Cutoff is the name of the concept. The implementation details described above are not directly related to the extension name. Sorry for the confusion caused by my poor explanation. Note. And this is the result when I confirmed the concept. I've developed |
@marson666 that's right. Enabling
|
Correct me if I am wrong, but according to the debug log shown by @marson666, the vector tables should be like
|
It would be nice if could give more details about the gather embeddings ? From the image from README.md, it seems that just simple concat the clip embeddings but still got an shape of 768*77, just same as the output from hide all tokens's.
The text was updated successfully, but these errors were encountered: