mstrazar/mklaren

A Python Multiple kernel learning library.

mklaren's Issues

align package issues

Trying to run from mklaren.regression.ridge import RidgeLowRank on python3.8 and getting ImportError: cannot import name 'Align' from 'align' error.

I believe updates are needed to comply with align package updates.

Where can I ge the align package?

Where can I ge the align package?
I'm using python 3.6

Gram matrices and unaligned data

Hi!
First of all I want to congratulate you for the amazing work put in this project, there are not so many stand alone MKL codes to use in Python, surprisingly.

I'm trying to use your work to apply multiple kernel regression over time-series. In case you don't know, this kind of data tends to be unaligned (not all the samples contain the same quantity of data). My form of preprocessing it is to create the Gram Matrix using several alignment kernel methods and then feed those matrices to the SVM.

I've tried several things. I'm using how alignment methods, multiplying the Gram Matrices of each kernel by the computed mu value. This, however, returns mediocre results. RidgeLowRank function only accepts Kinterfaces. I cannot create a Kinterface with the gram matrix directly, and inputing a numpy matrix with the original data of different row sizes makes the interface crash. KernelRidge suffers from the same problem. The Mklaren crashes using gram matrices. In 'cholesky_steps' function, a list called 'ina' empties before the loop ends and it crashes.

TL;DR

• RidgeLowRank only accepts Kinterfaces?
• A Kinterface can be build with numpy arrays of different row sizes?
• A Kinterface can be created with a gram matrix (with a precomputed kernel option, maybe)?
• Mklaren function seems to crash using gram matrices (tries to get argmax of a list that becomes empty)

Thank you so much for your patience.
~Pazaak

align package

Np, the issue is here. Where could I get the correct align package, that is used in mklaren/mklaren/mkl/l2krr.py ?

Pairwise constraints

I love the method and fantastic code base.

I'm wondering if you've thought about an extension for pairwise constraints rather than regression. For instance in alignf, the kernel alignment depends on the matrix Ky = y*y', where y is the predictor variable. Ky could rather be a matrix with entries k_ij, encoding side information about the relationship between sample i and j instead.

Also similar in setup to this method

thanks!

dependended module "kernel"

Hi

I am trying to install mklaren. But the package I installed is incomplete(automatically download from the mirrors.aliyun.com). Installing by the source is not available. So can you please update the pkg?

The details:
root@e1ace731c68a ~

pip install mklaren !493

Collecting mklaren
Downloading http://mirrors.aliyun.com/pypi/packages/08/d8/8adab90e7e5b85ccc46da94a20bcce1d397b0c85474a0eee515bfc4d4890/mklaren-1.0.tar.gz
Building wheels for collected packages: mklaren
Running setup.py bdist_wheel for mklaren ... done
Stored in directory: /root/.cache/pip/wheels/48/6a/9a/aed7546afe5300dbcd0fb7cfcf03d37ff592acb4b3e6bb5973
Successfully built mklaren
Installing collected packages: mklaren
Successfully installed mklaren-1.0
You are using pip version 8.1.2, however version 9.0.1 is available.
You should consider upgrading via the 'pip install --upgrade pip' command.

root@e1ace731c68a ~

python !494

Python 3.5.2 |Anaconda custom (64-bit)| (default, Jul 2 2016, 17:53:06)
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.

import mklaren
Traceback (most recent call last):
File "", line 1, in
File "/usr/anaconda3/lib/python3.5/site-packages/mklaren/init.py", line 1, in
import kernel
ImportError: No module named 'kernel'

More examples on how to optimize weights and choice of kernels

It would be great if the developers could comment on how to automatically choose optimal kernels (and their parameters) for a given set fo feature sets.. A grid search in an inner CV would obviously be the straightforward way to achive it but my be time consuming as well as unnecessary.

I have used Variational Bayed Probabilistic MKL from here before
https://www.cs.cornell.edu/~damoulas/Site/software.html

which helps optimizing the weights automatically, but choice of kernels stills needs to be made. Any thoughts? thanks.

Not updated documentation for poly_kernel

I've implemented 3 degree poly_kernel. But when I put kernel_args as 'p': 3, the program alerts an error saying that poly_kernel has no 'p' argument. So, I tried using 'degree' instead, and it worked.