Influence of batch size of training convergence · Issue #25 · openai/CLIP
github.com/openai/CLIP/issues/25
1 Users
0 Comments
3 Highlights
0 Notes
Tags
Top Highlights
he pairwise combinations of all image-text feature interactions are then cheap to compute on top of this, because they can re-use these already computed embeddings and only involve calculating a single additional inner product of relatively low-dimensional vectors per pair.
We had to modify BCE (manual downweighting to prevent negative pairs from dominating the loss) or "warm up" from a small amount of negatives.
Other common alternatives in the literature like triplet loss,
Ready to highlight and find good content?
Glasp is a social web highlighter that people can highlight and organize quotes and thoughts from the web, and access other like-minded people’s learning.