Implement MultiLabel losses and data input

[sergeyk]

Start with reference [1].

[1] Deep Convolutional Ranking for Multilabel Image Annotation
Yunchao Gong, Yangqing Jia, Sergey Ioffe, Alexander Toshev, Thomas Leung
http://arxiv.org/abs/1312.4894

[shelhamer]

See also:

[2] J. Weston, S. Bengio, and N. Usunier. Wsabie: Scaling up to large vocabulary image annotation. In
Proceedings of the Twenty-Second international joint conference on Artificial Intelligence-Volume Volume Three, pages 2764–2770. AAAI Press, 2011.
http://research.google.com/pubs/pub37180.html

[3] A. Frome, G. S. Corrado, J. Shlens, S. Bengio, J. Dean, M. A. Ranzato, and T. Mikolov. Devise: A deep visual-semantic embedding model. In Advances in Neural Information Processing Systems 26, pages 2121–2129. 2013.
http://research.google.com/pubs/pub41473.html

[sergeyk]

Note: HDF5DataLayer in #147 reads in labels as matrices already.

[kloudkl]

This issue is opened to be an umbrella of #144 in which @leezl has agreed to contribute back the MultiLabelDataLayer and all the related changes to proto, io etc.

The "loss function appropriate to multilabel data needs to be implemented" that @leezl talked about probably means extending the existing loss layers to support multilabel data. The todo list includes but is not limited to the SoftmaxWithLossLayer that is used by reference [1].

Besides, the WARPLossLayer implementing the ranking loss function in reference [1][2][3] in #126 demands multilabel data. In fact, any ranking loss function inherently requires multilabel data. I will keep synced with the progresses on the data layer.

[sguada]

@sergeyk, @shelhamer @jeffdonahue @Yangqing I'm implementing some multi-label losses and accuracy. But I would like to have the following:
Three different labels {-1,0,1}, which means negative, ignore and positive. Should I use 0 or -1 for negative, and conversely -1 or 0 for ignoring that label?

[sergeyk]

I'd say 0 for ignoring

[rbgirshick]

I second that (0 for ignoring).

On Wed, May 21, 2014 at 6:18 PM, Sergey Karayev notifications@github.comwrote:

I'd say 0 for ignoring

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/149#issuecomment-43838076
.

http://www.cs.berkeley.edu/~rbg/

[shelhamer]

By multiple label, do you mean ranking loss or one-vs-all style "attribute"
classification? For one-vs-all, I sketched up a loss layer that takes a
loss type and makes a classifier + loss pair for each class.

If ranking, or other true multi-class loss, that'll be sweet to have!

Re: how to ignore, I vote for 0 too.

Le mercredi 21 mai 2014, Ross Girshick notifications@github.com a écrit :

I second that (0 for ignoring).

On Wed, May 21, 2014 at 6:18 PM, Sergey Karayev <notifications@github.comjavascript:_e(%7B%7D,'cvml','notifications@github.com');>wrote:

I'd say 0 for ignoring

—
Reply to this email directly or view it on GitHub<
https://github.com/BVLC/caffe/issues/149#issuecomment-43838076>
.

http://www.cs.berkeley.edu/~rbg/

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/149#issuecomment-43846403
.

[sguada]

So the people have spoken, 0 would be for ignoring, -1 for negative and 1 for positive.

@shelhamer right now I'm changing the window_data_layer to generate multiple labels and modifying the cross_entropy for the loss, and creating a new multi-label_accuracy. So not a ranking loss yet.

If you could share your loss layer with me I would appreciate it.

[bhack]

Is this something different from last @sguada plans in #523 or could be closed?

[sguada]

This can be close #149 follow up in #523.

On Monday, September 22, 2014, bhack <notifications@github.com
javascript:_e(%7B%7D,'cvml','notifications@github.com');> wrote:

Is this something different from last @sguada https://github.com/sguada
plans in #523 #523 or could be closed?

—
Reply to this email directly or view it on GitHub
#149 (comment).

Sergio