Multi label support#1380
Conversation
It assumes labels are a vector of {-1,0,1} with -1 for negative, 1 for positive and 0 ignore
Conflicts:
src/caffe/proto/caffe.proto
Conflicts: src/caffe/layers/multi_label_accuracy_layer.cpp src/caffe/test/test_multi_label_accuracy_layer.cpp Forgot to fill the precision and f1-score in top Conflicts: src/caffe/layers/multi_label_accuracy_layer.cpp
Original source: dae5300 by sguada The merge conflicts are too many, so the original code is adapted and inserted manually onto the current dev branch.
|
With @mtamburrano we have starter to test this also merging your indirect layer branch. Do you still want to use the same sigmoid cross entropy changes done in @sguada PR? |
|
@bhack I am going to finish the implementation of As for the @sguada PR, I have already cleaned up the core of it in this branch, minus the part that would be rendered obsolete by |
|
@bhack , @sguada ,recently i am trying to implement the work in Street View House Number (pag. 13 http://arxiv.org/pdf/1312.6082v4.pdf), and i have a question, how can we write a global loss, and back propogate the loss the the multiple softmax layer? thanks |
|
I've followed #523 to here, I'm wondering when will this be merged into dev or master? |
|
Same here, waiting for this to be merged. |
There was a problem hiding this comment.
should this assignment outside for loop?
|
@netheril96 can you provide a example net definition showing how to use multi-label loss layers? |
|
when i try to built this pr i get the following error |
|
@netheril96, has this been abandoned? |
|
@danielhauagge I think that core developers want to declare this solved with #1698. So I think that will be hard to have multi label support from text file like with this PR. |
|
I have succesfully implemented multilabel support along with a soft max loss function. Will plan to submit a PR after ICCV deadline. I don't know why multi-label doesn't have support, it being a potentially important feature. @netheril96 @bhack |
|
Yes, I abandoned this PR. |
|
@sayanghosh can you share your work on multilabel support? |
12 participants
This is an attempt to clean up and improve upon @sguada work in #523. It is a work in progress, not ready for merge yet.