Skip to content

Triplet loss#3663

Open
eli-osherovich wants to merge 2 commits into
BVLC:masterfrom
eli-osherovich:triplet-loss
Open

Triplet loss#3663
eli-osherovich wants to merge 2 commits into
BVLC:masterfrom
eli-osherovich:triplet-loss

Conversation

@eli-osherovich

Copy link
Copy Markdown

This PR implements the triplet loss layer.

Eli Osherovich added 2 commits February 11, 2016 09:07
Signed-off-by: Eli Osherovich <oeli@amazon.com>
Signed-off-by: Eli Osherovich <oeli@amazon.com>
@cbalint13

Copy link
Copy Markdown
Contributor

@shaibagon

Copy link
Copy Markdown
Member

I see there is very little documentation in the code on how this loss works. If someone wish to understand it better (s)he might find this SO thread useful.

@eli-osherovich, thank you for implementing this loss.

@eli-osherovich

Copy link
Copy Markdown
Author

@shelhamer Is there any interest in this PR? If so, I can add documentation and examples.

@afshindn

Copy link
Copy Markdown

@eli-osherovich Are you planning on adding the examples and documentations?

@BeSlower

Copy link
Copy Markdown

Could you add one example and more documentations? Your code is very helpful for me, but a little bit hard to go through all things.

@jibweb

jibweb commented Oct 18, 2016

Copy link
Copy Markdown

@eli-osherovich Nice code, one question though, maybe I'm missing something obvious, but why is the line 65 of triplet_loss_layer.cpp
const Dtype scale = top[0]->cpu_diff()[0] / bottom[0]->num()

If I understood correctly, top[0]->cpu_diff()[0] would be the loss value, so why use this and not just 2 as in the derivation process ?

@kwin-wang

kwin-wang commented Nov 29, 2016

Copy link
Copy Markdown

@jibweb top[0]->cpu_diff[0] is equal to the loss_weight of loss layer in training prototxt file, the default value is 1. it's useful to multi-task learning.

@PeterouZh

Copy link
Copy Markdown

Hi, I want to know why I can't find triplet_loss_layer files in caffe's code now. I'll very thanks if anyone can help me.

@shaibagon

Copy link
Copy Markdown
Member

@pengzhou93 you cannot find files, because this PR was not merged yet. If you want to use it (at your own risk) you can pull this branch.

@soulslicer

Copy link
Copy Markdown

How does this vary from:

#5019

This one looks more extensive

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

9 participants