lucidrains/DALLE2-pytorch

some interesting results

miganchuanbo opened this issue · 4 comments

miganchuanbo commented

Hi,

Thank you for creating this repository.

I have found some interesting results :
The data(both image_embed and text_embed) are being l2-normed before they are fed to the model(is that right?).
But when I fed the model with data that have not being l2-normed, the quality of the sampled embed goes really bad(especially the cos sim).
I am wondering why.

AIDevMonster commented

thx for your asking.
I'm busy now.
I will send you the reason 2 days later.

miganchuanbo commented

thx for your reply, take your time. Let me know if there’s anything I can do to help

AIDevMonster commented

sure

rajakolla2 commented

Hi Miganchuanbo,
Can you elaborate what you meant by "the data(both image_embed and text_embed) are being l2-normed". Do you mean that they are being normalised in such a way that their l2-norm becomes equal to 1?