Handle the multi-label task?

[hsiangsky]

Hi Chou,

It seems that the libdnn couldn't handle the multi-label task.
Would you like to add this new feature in the future?

[poweic]

Sounds like a good idea !!
I'll add it in the future release. Thanks : )

[supergrover]

I'm interested in this too. Have you already worked on this? If not, I forked your repository and am working on it right now.

[poweic]

That's great !! I haven't started yet.

Here's my idea. It's rough and maybe you can help me with this.
Originally, the label comes along with the feature. Like this:

12 1:1 2:0.5 10:0.7 ...

where 12 is the label, and the rest 1:1 2:0.5 10:0.7 ... represent the feature.

For multi-labeled feature, I was thinking about providing another label file. Like this:

# feature file
1:1 2:0.5 10:0.7 ...

where 12 is missing and 1:1 2:0.5 10:0.7 ... still represent the feature.

# label file
12 15 17

where the above feature not only labeled 12, but also 15 and 17.

Besides, nn-train and class BasicStream should tell the difference between them and ask user to provide an additional label file if it's multi-labeled.

Any suggestion?

[supergrover]

That was pretty much what I had in mind. I will work out some stuff (still reading in on your code), and will come back on this.

[poweic]

Hi~ @supergrover, @hsiangsky
Have you guys started yet ?

I almost finish the support of multi-label.
Lots of refactoring in src/data-io.cpp, src/dataset.cpp, include/data-io.h and include/dataset.h.
Guess it's going to be cleaner.

[supergrover]

@botonchou working on it now

[poweic]

Because my data IO class sucks, I decided to refactor it first.
it's on another unpublished branch. (it's more readable and cleaner now)

Most of the functions needed in multi-class classification are done.
(Except for the measurement of multiclass accuracy.)
I'm going to publish it in about a week.

Do you want to start from that branch ?

[supergrover]

Yes, please push. I will play around with it, although for consistency I think you should finish it

[poweic]

It's on branch feature/multi-label. Thanks : )