Hi,
Just noticed that I coincidentally sent half of my email!! Anyway, here is
my question again:
A couple of weeks ago I test the intersection kernel in my experiments, for
that I wrote the following simple procedures:
import numpy as np
import cv2
def intersection_dist(v1,v2):
return
2012/9/20 abdalrahman eweiwi :
> Hi,
>
> Just noticed that I coincidentally sent half of my email!! Anyway, here is
> my question again:
>
> A couple of weeks ago I test the intersection kernel in my experiments, for
> that I wrote the following simple procedures:
>
> import numpy as np
> import c
On Thu, Sep 20, 2012 at 10:54 AM, Olivier Grisel
wrote:
> 2012/9/20 abdalrahman eweiwi :
> > Hi,
> >
> > Just noticed that I coincidentally sent half of my email!! Anyway, here
> is
> > my question again:
> >
> > A couple of weeks ago I test the intersection kernel in my experiments,
> for
> > th
You can also debug by printing the kernel value of pairs of samples
that you know are qualitatively related and pairs that you know are
qualitatively unrelated.
If you kernel is positive (it should be), pairs of related samples
should always have a higher value than pairs of unrelated samples.
--
*
*
*return np.min(np.vstack((v[None,:],v2[None,:])),axis=0).sum(*)
I'm not sure what the None does here, I would write
np.minimum(v, v2).sum()
which is the correct intersection kernel I think.
What are the features you use? The features have to be positive for this
to work.
There should be
On Thu, Sep 20, 2012 at 12:58:56PM +0100, Andreas Mueller wrote:
> return np.min(np.vstack((v[None,:],v2[None,:])),axis=0).sum()
>I'm not sure what the None does here, I would write
None create a new axis. For readability, one should always write
v[np.newaxis, :], which is equivalent.
On 09/20/2012 01:01 PM, Gael Varoquaux wrote:
> On Thu, Sep 20, 2012 at 12:58:56PM +0100, Andreas Mueller wrote:
>> return np.min(np.vstack((v[None,:],v2[None,:])),axis=0).sum()
>> I'm not sure what the None does here, I would write
> None create a new axis. For readability, one should al
On 09/20/2012 01:01 PM, Gael Varoquaux wrote:
> On Thu, Sep 20, 2012 at 12:58:56PM +0100, Andreas Mueller wrote:
>> return np.min(np.vstack((v[None,:],v2[None,:])),axis=0).sum()
>> I'm not sure what the None does here, I would write
> None create a new axis. For readability, one should al
On Thu, Sep 20, 2012 at 2:08 PM, Andreas Mueller
wrote:
> On 09/20/2012 01:01 PM, Gael Varoquaux wrote:
> > On Thu, Sep 20, 2012 at 12:58:56PM +0100, Andreas Mueller wrote:
> >> return np.min(np.vstack((v[None,:],v2[None,:])),axis=0).sum()
> >> I'm not sure what the None does here, I wou
On Thu, Sep 20, 2012 at 1:58 PM, Andreas Mueller
wrote:
>
> *
> *
> *return np.min(np.vstack((v[None,:],v2[None,:])),axis=0).sum(*)
>
> I'm not sure what the None does here, I would write
> np.minimum(v, v2).sum()
> which is the correct intersection kernel I think.
>
I know, sometimes i make thi
2012/9/20 abdalrahman eweiwi :
>
>
> On Thu, Sep 20, 2012 at 1:58 PM, Andreas Mueller
> wrote:
>>
>>
>>
>> return np.min(np.vstack((v[None,:],v2[None,:])),axis=0).sum()
>>
>> I'm not sure what the None does here, I would write
>> np.minimum(v, v2).sum()
>> which is the correct intersection kernel
>>> There should be no other parameters to tune here.
>>>
>>> Btw, you can try kernel_approximation.AdditiveChi2Sampler (if your data is
>>> positive).
> Note that this is not meant to be used as an exact kernel in a kernel
> SVM but as a explicit preprocessing transformer for a linear kernel.
> T
Hello,
I was asking Olivier about CRF in sklearn and I ended up discussing some of
my Kaggle experience.
I am forwarding my email to this list on his suggestion.
Thanks to the sklearn team (especially the text classification module
authors) for helping me win a competition at Kaggle (it was a pri
Hey Vivek.
Thanks for sharing your success with sklearn.
Out of curiosity: CRF learning for loopy graphs or non-loopy graphs or
chains?
On 09/20/2012 05:17 PM, Vivek Sharma wrote:
>
>
> Features that I wish were available:
> - openmp/multiprocessing mode for libsvm
> (http://www.csie.ntu.edu.t
Hi Andreas,
Thanks for your comments.
For CRF, linear chain... mainly for NLP tasks.. I will probably use the
Stanford implementation, but I haven't tried it yet.
I agree that sparse RF/GBM may not be all that useful, because if the
problem is too sparse, its probably better to use something els
2012/9/20 Vivek Sharma :
> Hi Andreas,
>
> Thanks for your comments.
>
> For CRF, linear chain... mainly for NLP tasks.. I will probably use the
> Stanford implementation, but I haven't tried it yet.
>
> I agree that sparse RF/GBM may not be all that useful, because if the
> problem is too sparse,
Hah, I saw the same paper and when I saw it was an option in sklearn, I
thought it must be a good idea. I couldn't make it work though - it worked
under some condition, but not in others. I don't remember the exact issue.
On Thu, Sep 20, 2012 at 2:05 PM, Peter Prettenhofer <
peter.prettenho...@gm
Below are some excerpts from the "Build failed" message that I got
after git rm'ing the sparse linear models code. The strange thing is
that it seems to start rebuilding in the middle of the tests. The same
thing happened when I tried nosetests sklearn/tests/ on my own box,
which also produced a lo
On Mon, Sep 17, 2012 at 9:05 AM, Jacob VanderPlas
wrote:
> Our plan is to propose two tutorials: I'll submit an intro to machine
> learning with scikit-learn, based on the material I covered at Scipy
> 2012 (http://astroML.github.com/sklearn_tutorial). Olivier will submit
> a more advanced tutori
ps - to match things, we'll mention in ours that it would be good to
schedule it as early as possible, so that any other tutorials on
sklearn, pandas, numpy, etc, can come later and attendees who take
ours can benefit from the combination. The pycon guys actually care
about that stuff, so I'm sure
Hi,
On Thu, Sep 20, 2012 at 12:17:07PM -0400, Vivek Sharma wrote:
> Thanks to the sklearn team (especially the text classification module
> authors) for helping me win a competition at Kaggle (it was a private
> competition, organized in April).
Well, thanks for saying thanks! :)
It is a natural
21 matches
Mail list logo
