-
Notifications
You must be signed in to change notification settings - Fork 66
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
feat: integrate new loss and pooling options #664
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Great work! I like it!
i feel the name |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nice documentation, added some comments.
1b10bb4
to
98c6d5c
Compare
the documentation is not included in the docs anymore, please double check |
Ah, I forgot to update the index when I renamed the file, fixed now |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM! If the broken link is fixed.
|
||
SphereFace loss is a loss function that was first formulated for computer vision and face recognition tasks. | ||
Finetuner supports two variations of this loss function, `ArcFaceLoss` and `CosFaceLoss`. | ||
Instead of attempting to minimise the distance between positive pairs and maximise the distance between negative pairs, the SphereFace loss functions compare each sample with the center of each class |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Instead of attempting to minimise the distance between positive pairs and maximise the distance between negative pairs, the SphereFace loss functions compare each sample with the center of each class | |
Instead of attempting to minimise the distance between positive pairs and maximise the distance between negative pairs, the SphereFace loss functions compare each sample with the center of each class |
I think here it is confusing what is the center of a class ?
In practice this is not really center but more estimated prototypes which is learn alongside the representation in the last linear softmax layer. Not sure how to put it in nice word so I let you the burden
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I have changed it to '...with an estimate of the center point of each classes' embeddings.'
I don't want to go into too much detail explaining it but i think this is more accurate
1e9a40d
to
db4ce5f
Compare
db4ce5f
to
01b3fb2
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Add a few more comments
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM!
📝 Docs are deployed on https://ft-feat-support-arcface--jina-docs.netlify.app 🎉 |
This pr adds support for the newly added loss and pooling options from core, as well as a new page in the documentation explaining how to use them.