I would definitely join the sprint, anything after June 17 works for
me. I was thinking to come hang around during ICML, even if I might
not be able to afford the conference.
Cheers,
Vlad
On Tue, Apr 12, 2016 at 11:39 AM, Andreas Mueller wrote:
> So should we pick another or
I usually use an absolute threshold for min_df and a relative one for
max_df. I find it very useful to look at the histogram of word dfs for
choosing the latter, it varies a lot from dataset to dataset. For
short texts, like tweets, words such as "the" can have a df of 0.1.
It's very easy to look
Hi Mamun,
If your cluster labels are known, you can use the LabelShuffleSplit
ore LeavePLabelOut cross-validation generators.
HTH,
Vlad
On Fri, Feb 5, 2016 at 10:05 AM, Mamun Rashid wrote:
> Hi Folks,
> I have a two class classification problem where the positive
Hi James,
I'm not sure how useful a minimum alpha would be. Even if no weights
are shrunk quite to zero, the regularization can still impact
performance metrics. I would be curious what application you have in
mind for this.
The max alpha question is interesting, I am curious as well. (Sorry my
In the case of "char_wb" it sounds indeed like a custom tokenizer
should be called if given. That would require a different
implementation than the current one, however. You might want to file
an issue.
Sebastian's suggestion works, but note that scikit-learn's default
tokenization is not the
Is there a reason why you are (still) not respecting the API constraints for
custom estimators given in the documentation?
__init__ should only set parameters on self that have (exactly) the same name
as the arguments passed to it.
Your __init__ should be:
self.k = k
Another thing I've seen people do is to threshold based on the
difference between the scores of the best and second best topics.
(Only take documents with a clear winning topic.) For estimating the
number of topics, you can use cross-validation.
Vlad
On Wed, Apr 29, 2015 at 12:42 AM, Joel
Hi Roberto
what does None do for max_depth?
Copy-pasted from
http://scikit-learn.org/stable/modules/generated/sklearn.tree.DecisionTreeClassifier.html
If None, then nodes are expanded until all leaves are pure or until all leaves
contain less than min_samples_split samples.”
In particular,
, Pagliari, Roberto wrote:
Hi Vlad,
when using randomized grid search, does sklearn look into intermediate
values, or does it samples from the values provided in the parameter grid?
Thank you,
From: Vlad Niculae [zephy...@gmail.com]
Sent: Monday, April 20
-optimization
Vlad
On 20 Apr 2015, at 15:34, Vlad Niculae zephy...@gmail.com wrote:
The example you cite contains these lines:
max_features: sp_randint(1, 11),
min_samples_split: sp_randint(1, 11),
min_samples_leaf: sp_randint(1, 11),
Those
In order to support discrete parameters, our tree implementation would need
to support categorical variables though.
Ah, good point, I didn’t think about that. But we could use the usual hacks
(integer or one-hot encoding). I wonder how that compares to using GPs and
rounding when it
Hi Gael,
On 31 Mar 2015, at 14:01, Gael Varoquaux gael.varoqu...@normalesup.org
wrote:
Why do you think the GP route is easier?
Because we already have GPs.
Well, we already have random forests too.
Both cases would need quite a bit of machinery on top, and I don’t know the
extent of
Hi Boyuan, hi everyone,
On top of what Andy said, I would like to add that you don’t have to commit to
certain algorithms in the proposal, as long as you make the plan very clear,
and you leave time for discussing alternatives, pros and cons with the
community.
Since you say there is some
Hi Raghav, hi everyone,
If I may, I have a very high-level comment on your proposal. It clearly shows
that you are very involved in the project and understand the internals well.
However, I feel like it’s written from a way too technical perspective. Your
proposal contains implementation
Hi Wei Xue, hi everyone,
I think Andy’s comments about testing and documentation are very important.
I have just a few things to add:
1. As confused as I am about the world around me, I still knew that the current
year is 2015 :P I think that the form is asking “which year of your program you
Hi Cristoph, Gael, hi everyone,
On 24 Mar 2015, at 18:09, Gael Varoquaux gael.varoqu...@normalesup.org
wrote:
Don't you think that I could also benchmark models that are not
implemented in sklearn? […]
I am personally less interested in that. We have already a lot in
scikit-learn and
Hi Vinayak,
The wiki page just lists a subset of possible topics for which candidates
already showed concrete interest. I think an application for low-rank matrix
completion would be more than welcome. It’s very important to work on a topic
that you are interested in directly, versus just
some work upon that, but I didn't
get any feedback.
On Tue, Mar 24, 2015 at 3:23 AM, Vlad Niculae zephy...@gmail.com wrote:
Hi Vinayak,
The wiki page just lists a subset of possible topics for which candidates
already showed concrete interest. I think an application for low-rank matrix
Apologies in advance, but this fits so well, I couldn’t help myself.
A Mathematician and an Engineer attend a lecture by a Physicist. The topic
concerns Kulza-Klein theories involving physical processes that occur in spaces
with dimensions of 9, 12 and even higher. The Mathematician is sitting,
Hi Roberto,
This is explained in the Python standard library documentation:
https://docs.python.org/3/library/functions.html#sorted
Cheers,
Vlad
On 18 Feb 2015, at 21:33, Pagliari, Roberto rpagli...@appcomsci.com wrote:
what does sorted do if the best average cv score is the same?
how
Hi Roberto,
This is all documented in more detail here: [1]
The transform looks good (just that you might want to add a flag to avoid
memory copies when you can afford to destroy the original data).
It’s not clear what the intention of `my_param` is here. It’s not user
specified, right?
Hi Roberto,
Everything I say below is also explained in the developers documentation that I
linked to in the other e-mail. [1]
You are breaking some conventions that make the default `get_params` and
`set_params` not work well.
As I said in the other thread, fitted attributes are suffixed
On 11 Feb 2015, at 16:31, Andy t3k...@gmail.com wrote:
On 02/11/2015 04:22 PM, Timothy Vivian-Griffiths wrote:
Hi Gilles,
Thank you so much for clearing this up for me. So, am I right in thinking
that the feature selection is carried for every CV-fold, and then once the
best
Hi Luca
x_3_dimensional = x.dot(spca.components_.T) # this is equivalent to
spca.transform(x)
This part is specific to PCA. In general, the transform part of such a
decomposition is `X * components ^ -1`. In PCA, because `components`
is orthogonal, `components ^ -1` is `components.T`. The
To clarify, it is *not* the case that `x.dot(spca.components_.T) ` is
equivalent to `spca.transform(x)`. The latter performs a solve.
Best,
Vlad
On Fri, Oct 17, 2014 at 12:03 PM, Vlad Niculae zephy...@gmail.com wrote:
Hi Luca
x_3_dimensional = x.dot(spca.components_.T) # this is equivalent
Hi Luca,
The other part of the decomposition that you're missing is available
in `spca.components_` and has shape `(n_components, n_features)`. The
approximation of X is therefore `np.dot(x_3_dimensional,
spca.components_)`.
Best,
Vlad
On Thu, Oct 16, 2014 at 6:07 PM, Luca Puggini
Hi Zoraida,
The Imputer assumes that your data is a numeric numpy array, or
convertible to one. You should replace your string NA values with
np.nan objects, then use the Imputer with the default,
`missing_values='NaN'`.
It's easier to debug if you explicitly convert your data to a float
numpy
Hi Pavel,
First of all, this is an interesting subject, thanks for bringing it
up! I fear that it's too domain-specific to go very deep in this
direction.
That being said, and trying to interpret your benchmarks, it seems
that Delta-idf might actually be interesting.
Or, more generally, the idea
It has confused me as well, +1.
It's counterintuitive and broken, in my opinion.
Vlad
On Wed, Aug 20, 2014 at 2:31 PM, Gael Varoquaux
gael.varoqu...@normalesup.org wrote:
It's been around for so long, but it's also hard to believe that anyone
exploited this behaviour intentionally. Shall we
Also, the class is well documented, but because of an omission, it
wasn't linked from the API page at the time of the last stable
release.
This has been fixed in the development version, so you can read the
docs in a friendlier way here [1].
Best,
Vlad
[1]
Hi,
If you want to get `sample_weights` working with the current master,
the easiest is to take PR 3524 and either pass it through `fit_params`
or just undo the last commit in the branch.
I needed to change a couple of things to get 1574 up to date with the
current master, but nothing else is
Hi,
Allow me to clarify. We don't implement Hoyer's sparse update rule
indeed (it shouldn't say this implements, I initially cited Hoyer
for motivating sparseness constraints in NMF). Instead, we implement a
version of sparse NMF with a clear (but not particularly elegant)
objective function,
wrote:
i will never post a docstring again :)
sorry for the noise
michael
On Wednesday, June 25, 2014, Vlad Niculae zephy...@gmail.com wrote:
Hi,
Allow me to clarify. We don't implement Hoyer's sparse update rule
indeed (it shouldn't say this implements, I initially cited Hoyer
IIRC, weekly post are not a GSoC requirement but they are a _PSF_
requirement, and since scikit-learn is participating to GSoC under the
PSF umbrella, the requirement applies to us. I think it's great
incentive to think of your work in terms of what you could show to
others. No matter how
This is great news, congratulations Gilles!
Cheers,
Vlad
On May 22, 2014 8:15 AM, Gilles Louppe g.lou...@gmail.com wrote:
Hi folks,
Just for letting you know, my talk Accelerating Random Forests in
Scikit-Learn was approved for EuroScipy'14. Details can be found at
Hi John,
I believe general inference methods are out of scope for scikit-learn.
Even general structured learning algorithms are not in scope at the
moment, as it's hard to fit problems in numpy arrays. For learning,
you might want to check out pystruct [1].
If you just want inference,
Hi Manoj,
For efficiency, the BLAS api defines different functions for different
underlying datatypes (float32, float64, complex64, complex128). The
scipy get_blas_funcs utility has the role of getting the Python
wrapper for the given BLAS functions (in this case 'swap' and 'nrm2',
that's
This program is granted free of charge for research and education
purposes. However you must obtain a license from the author to use it
for commercial purposes.
Definitely FEST is not BSD compatible :(
Vlad
On 17/3/2014 14:19 , Arnaud Joly wrote:
Hi,
The support for sparse matrices
In some cases it might be preferable to fit an OvA model. In those
cases, I think the
user code would look nicer and more explicit if it'd use the
sklearn.multiclass.OneVsRest encoder.
The downside is that we'll need to go through an ugly deprecation cycle
for a major class in the library.
On Wed Feb 26 13:32:08 2014, Gael Varoquaux wrote:
documentation and example
This was exactly my thought. Many such (near-)equivalences are not
obvious, especially
for beginners. If Lars's hinge ELM and RBF network would work well (or
provide
interesting feature visualisations) on some
If you're affiliated with a university, Anaconda has free academic
licenses that include MKL and their optimized builds.
Vlad
On Mon Feb 24 09:22:07 2014, Javier Martínez-López wrote:
That is great, thanks! I do not have the mkl module (it isn't free,
right?) but with your script the
Hi Manoj,
In the first example, the intercept is not regularized, hence the
difference.
Vlad
On Feb 15, 2014 8:54 AM, Manoj Kumar manojkumarsivaraj...@gmail.com
wrote:
Hello
I have a query with fit_intercept parameter in most of the estimators.
When we have a linear model like w0 + w1*x1 +
I've heard stchee-kit once, along with stchee-pee and num-pee.
Vlad
On Sun Feb 2 18:39:58 2014, Hadayat Seddiqi wrote:
i always said skikit
On Sun, Feb 2, 2014 at 12:20 PM, Andy t3k...@gmail.com
mailto:t3k...@gmail.com wrote:
On 02/02/2014 12:06 PM, Olivier Grisel wrote:
Note:
I like the locality-sensitive hashing idea!
Vlad
On Tue Jan 28 10:04:36 2014, Nick Pentreath wrote:
This would be a great addition.
Some ideas /code perhaps: http://nearpy.io/
On Tue, Jan 28, 2014 at 10:59 AM, Mathieu Blondel
math...@mblondel.org mailto:math...@mblondel.org wrote:
I don't think Weka (at least the interesting parts of it) could run on
Android either. I don't really foresee the whole Scipy stack running on
Android; maybe one day when all dependencies are rewritten in PyPy and
are faster and still 100% compatible...
One thing that would be possible (but I
Hi Arnaud, awesome poster! Here are a few things that popped out:
Firstly, I doubt it matters, but some of the links are mangled.
Then, I think it should say students' master's theses or something
like this (plural). Also the chromosome 15 sounds strange to me
compared to chromosome 15.
I would rather have this sorted out through the github issue tracker.
I don't think it's a good idea to encourage users to e-mail individual
developers. Someone else could have the expertise and do the change
confidently.
My 2c,
Vlad
On Thu Jan 16 18:12:05 2014, Issam wrote:
Hi scikit-learn
Works exactly as you described on my machine (which doesn't mean much
because it's relatively close to yours, but I am just too enthusiastic
about this not to reply! \o/)
Memory usage is as expected. I see a speedup in train time but a
slight slowdown in test time (1.7 vs 1.0), is it expected or
propose to turn off multiprocessing at prediction time - this
might backfire quite easily.
2013/12/20 Olivier Grisel olivier.gri...@ensta.org
2013/12/20 Vlad Niculae zephy...@gmail.com:
Works exactly as you described on my machine (which doesn't mean much
because it's relatively close to yours
Great, thanks a lot!
I'm also curious about what you're running it on and about how the
performance is.
Vlad
On Fri, Dec 13, 2013 at 7:11 PM, Olivier Grisel
olivier.gri...@ensta.org wrote:
Nice.
Have you used it with success for real image classification tasks?
I see you have been involved
On Fri, Dec 13, 2013 at 12:20 PM, Vlad Niculae zephy...@gmail.com wrote:
Great, thanks a lot!
I'm also curious about what you're running it on and about how the
performance is.
Vlad
On Fri, Dec 13, 2013 at 7:11 PM, Olivier Grisel
olivier.gri...@ensta.org wrote:
Nice.
Have you used
Personally I'd rather be a bit frustrated but have tab completion and
pyflakes warnings. I avoid using star imports even in hackish scripts.
I assume the warning will create unnecessary confusion when people
learn to use the star import first. These users will probably feel
that the warning is a
I guess remove means deprecate, right?
I am +1 but we should definitely find a place for the code. Worse case
it will be a repo with containing just the HMM.
My thoughts exactly; my impression is that people do find the code
useful and it's reasonably readable. It should definitely go into a
seqlearn uses a different API on purpose though (one big ndarray),
whereas pystruct uses lists of arrays but is only focused on
max-margin learning :)
On Sat, Nov 30, 2013 at 12:38 PM, Gael Varoquaux
gael.varoqu...@normalesup.org wrote:
+1 on the whole thread.
I was hoping that Lars's seqlearn
it will be the same one. But I'm not the
best person to ask, I've never even used the Kernel PCA.
Cheers,
Vlad
-- Forwarded message --
From: Vlad Niculae v...@vene.ro
Date: Mon, Nov 25, 2013 at 10:41 PM
Subject: Fwd: Problem with scikit learn kernel PCA
To: Vlad Niculae zephy...@gmail.com
I finally found a desk and some focus. I addressed Mathieu's
suggestions and added some timings on real data (with a lot of
concessions so that it would run reasonably quick on my machine).
Here's the results: http://nbviewer.ipython.org/7224672
It becomes clear that `tol` still means different
Re: the discussion we had at PyCon.fr, I noticed that the internal
elastic net coordinate descent functions are parametrized with
`l1_reg` and `l2_reg`, but the exposed classes and functions have
`alpha` and `l1_ratio`. Only yesterday there was somebody on IRC who
couldn't match Ridge with
We have an instance of vbench continuously running [1] that I did as a
GSoC project last year.
For some reason it seems that the links don't generate properly now,
but it still works (though all data got lost in a jenkins setup
incident this summer).
Here are some linear model benchmarks for
Vlad, that's exactly what I've been looking for!
Thanks,
Karol
2013/11/8 Vlad Niculae zephy...@gmail.com
We have an instance of vbench continuously running [1] that I did as a
GSoC project last year.
For some reason it seems that the links don't generate properly now,
but it still works
, Olivier Grisel olivier.gri...@ensta.org wrote:
2013/11/7 Vlad Niculae zephy...@gmail.com:
Hi everybody,
I just updated the gist quite a lot, please take a look:
http://nbviewer.ipython.org/7224672
I'll go to sleep and interpret it with a fresh eye tomorrow, but
what's interesting at the moment
?
On Thu, Nov 7, 2013 at 11:12 AM, Vlad Niculae zephy...@gmail.com wrote:
The regularization is the same, I think the higher residuals come from
the fact that the gradient is raveled, so compared to `n_targets`
independent problems, it will take different steps.
I don't think there are any
4 35.7 MiB 0.0 MiB def linalg(X):
5 42.7 MiB 7.0 MiB return np.linalg.norm(X, 'fro')
On Thu, Nov 7, 2013 at 11:46 AM, Vlad Niculae zephy...@gmail.com wrote:
Come to think of it, Olivier, what do you mean when you say L-BFGS-B
has higher residuals
0.9300757
2900.9297058
2970.9262745
3040.9274619
3110.9275654
Name: residual, dtype: object
It looks spot on. Note that tolerance is 1e-3. Any idea how to make
it visible in the plot when two lines are so close?
On Thu, Nov 7, 2013 at 12:26 PM, Vlad Niculae zephy...@gmail.com wrote
This is a known problem with np.linalg.norm, and so is the memory
consumption. You should use sklearn.utils.extmath.norm for the
Frobenius norm.
Hmm. Indeed I missed that, but still, this is a bit odd.
sklearn.utils.extmath.norm is slower than raveling on my anaconda with
MKL accelerate setup:
I feel like this would go against explicit is better than implicit,
but without it grid search would indeed be awkward. Maybe:
if self.alpha_coef == 'same':
alpha_coef = self.alpha_comp
?
On Thu, Nov 7, 2013 at 4:19 PM, Mathieu Blondel math...@mblondel.org wrote:
On Thu, Nov 7, 2013 at
lasso (as well as the sparse variant). Is there any other reason for
this or just that nobody needed it?
Cheers,
Vlad
On Wed, Oct 30, 2013 at 10:40 AM, Vlad Niculae zephy...@gmail.com wrote:
Thanks Mathieu, well part of it comes from your gist (I added an
attribution now) ;)
Non-negative
i guess it's just a bug in how the solvers return residuals, I'll add
some unit tests with manually-computed residuals to check.
On Wed, Oct 30, 2013 at 9:48 AM, Olivier Grisel
olivier.gri...@ensta.org wrote:
Does anyone have a explanation for the discrepancy in the residuals
for the lbfgs-b
Thanks Mathieu, well part of it comes from your gist (I added an
attribution now) ;)
Non-negative lasso is really interesting, I forgot about it but I
think it would be very interesting to compare qualitatively.
Vlad
On Wed, Oct 30, 2013 at 10:15 AM, Olivier Grisel
olivier.gri...@ensta.org
Hi all,
During the PyCon sprint I kept digging into the NMF and specifically
ways to solve each sub-iteration. It became clear that the alternating
NLS approach finds good reconstructions and converges well, but the
NLS solving step is critical and must be optimized.
I have started looking into
Hi,
We refer to such a setting as *multi-label*. Please take a look at
http://scikit-learn.org/stable/modules/multiclass.html
Yours,
Vlad
On Sun, Oct 20, 2013 at 1:19 PM, Mahendra Kariya
geek3142-skle...@yahoo.co.in wrote:
Hi,
I am trying to do multi class classification using NB or linear
There are still a few things that are not clear to me from the
documentation. Can you customize the classifier to perform a different
decision function?
You can subclass it and override the decision_function method.
While true, this can be misleading. You're just changing the final
step used
Just to add, I don't think you need to reshape y. And reshaping x can be
more briefly stated as x[:, np.newaxis].
In my opinion supporting such cases, while convenient for users, would lead
to annyoing branches and code that is harder to maintain and test. The
important thing is being consistent.
And under the current implementation, implementing them involves
changing only the sampling and energy computation, I think.
I discussed this with Gabriel Synnaeve during the sprint and I think
he was working on the gaussian version, it might be on his repo.
Lars, do you have any practical
Also, the builds fail quite rarely (with the exception of the last few
weeks). And when they do, I think these e-mails make sure that it gets
fixed faster than without them. It's better not to unsubscribe. Even
if it's annoying if it's *definitely* not your fault (documentation
PRs) sometimes you
PM, Olivier Grisel
olivier.gri...@ensta.orgwrote:
2013/8/28 Lars Buitinck l.j.buiti...@uva.nl:
2013/8/28 Vlad Niculae zephy...@gmail.com:
Do the indices/indptr arrays need to be int32 or is this a limitation
of the
implementation?
This is a limit in scipy.sparse, which uses signed int
It's about redirecting /dev and /stable to the appropriate fixed paths.
Actually I remember that this has been looked into, I vaguely remember a
thread a while back.
I think the problem is that we couldn't move to github while keeping all
the old links and looking the same in the eyes of the
If you're writing an external script that just interfaces with
scikit-learn and you intend to keep it separately distributable (3rd
party), you can replace them with absolute imports:
```
from sklearn.base import ClassifierMixin, RegressorMixin
from sklearn.externals.joblib import Parallel,
I'll have to side slightly against Lars on this one.
I agree with Lars that any software that doesn't support these is broken,
that Unicode looks better than other ad-hoc formatting.
If the software works, often the fonts won't. Personally if I'd need to
see the source and find characters
Hi all,
I got an unexpected error with current master, when trying to run
TfidfVectorizer on a 2 billion token corpus.
/home/vniculae/envs/sklearn/local/lib/python2.7/site-packages/sklearn/feature_extraction/text.pyc
in _count_vocab(self,
raw_documents, fixed_vocab)
728
After doing it again with pdb I figured out that it has nothing to do with
vocabulary size, which is decent; the list of indices simply grows too big.
Vlad
On Wed, Aug 28, 2013 at 11:01 PM, Vlad Niculae zephy...@gmail.com wrote:
Hi all,
I got an unexpected error with current master, when
on nosetests. I have made some progress using
scikits and learning python but I never got that to work.
Thanks again,
Don
On Aug 24, 2013, at 6:16 PM, Vlad Niculae zephy...@gmail.com wrote:
The `python` and `nosetests` executables that you are running are probably
not the macports ones. Type `which
Is it maybe related to the OS, as it seems that the problem is with opening
the memmapped file?
Vlad
On Sat, Aug 24, 2013 at 1:52 PM, Olivier Grisel olivier.gri...@ensta.orgwrote:
Sounds like a serious bug, could you please open an issue on github?
--
Olivier
Congratulations Andy! Thanks for all your hard work on this.
This is a good moment for pystruct to gain some momentum!
Cheers,
Vlad
On Sun, Aug 11, 2013 at 8:55 PM, Andreas Mueller
amuel...@ais.uni-bonn.de wrote:
Hey everybody.
I just wanted to spam the ML again and say I just released
Sorry, but I can't find the issue, you posted the same link twice.
Those errors are very similar to what I was getting before figuring
out that I need to use nosetests3 instead of nosetests.
Vlad
On Mon, Jul 29, 2013 at 10:35 AM, Olivier Grisel
olivier.gri...@ensta.org wrote:
I found problems
I can do it; the question is whether to build against anaconda or
against binary numpy/scipy; and whether it matters. I'll see if I can
check.
On Mon, Jul 29, 2013 at 12:09 PM, Olivier Grisel
olivier.gri...@ensta.org wrote:
2013/7/29 Olivier Grisel olivier.gri...@ensta.org:
I found problems
Or simply hide the 0.14a1 release? It should still stay pip
installable if you use the right magic words, right?
On Mon, Jul 29, 2013 at 1:35 PM, Andreas Mueller
amuel...@ais.uni-bonn.de wrote:
On 07/29/2013 01:20 PM, Andreas Mueller wrote:
On 07/29/2013 01:13 PM, Olivier Grisel wrote:
Maybe
I uploaded the windows binaries manually through the web interface
with no issue.
Unrelated question: We could go for a python3.3 binary too, but I
would need to build it using the (free) scipy installed with Anaconda,
because official scipy doesn't provide binaries for python 3.3. From
what I
/
On Mon, Jul 29, 2013 at 1:58 PM, Gael Varoquaux
gael.varoqu...@normalesup.org wrote:
On Mon, Jul 29, 2013 at 01:54:21PM +0200, Vlad Niculae wrote:
I uploaded the windows binaries manually through the web interface
with no issue.
I might give up and upload it manually, but I tend to like
Hi Harold,
Only the current development version, and the upcoming release, has,
as of recently, support for Python 3. Even so, it won't be easy to
support 3.2, we just aim for 3.3 at the moment.
This being said, I have no idea what causes this specific error. That
line seems unchanged in the
The unable to find vcvarsall.bat error is because you don't have
environment variables set appropriately. Click StartProgramsVisual
Studio C++ Express Visual Studio Command Prompt and run the setup
from there.
Vlad
On Mon, Jul 22, 2013 at 11:08 AM, Andreas Mueller
amuel...@ais.uni-bonn.de
.
The problem might be the unavailability of blas implementation in Windows as
I figured out.
Numpy doesn't have settings for blas.
In Linux versions we need to get dependencies for blas (libatlas-dev) before
builiding. But in Windows it's not there.
On Mon, Jul 22, 2013 at 2:44 PM, Vlad
Also depending on the model you want to deploy, if you just need to
predict using a pre-trained model you can extract the decision
function and the data of the model and rewrite it in another language.
In many cases applying a trained model is very easy.
Vlad
On Wed, Jul 17, 2013 at 12:31 AM,
The requirements are definitely the blocking thing here. Not just the
dependency on cvxopt but also the inference packages and the fact they
need to be built manually. The api is sklearn-ish enough even with
lists-of-lists.
On Fri, Jul 12, 2013 at 10:06 AM, Andreas Mueller
If you have MSVC from C++ express 2008 available could you try with that?
Are you trying to build the latest master, does the last release work well?
Vlad
On Thu, Jul 11, 2013 at 5:17 PM, Maheshakya Wijewardena
pmaheshak...@gmail.com wrote:
I do not have MKL.
Can there be any other reason for
Hi Mathieu,
Will you be joining online? People have been asking this on IRC ;)
Personally I want to take care of unfinished business like the omp CV,
the RBM pull request, GSOC PRs, and I was thinking of trying to tackle
Averaged SGD; apart from this I'll be side-sprinting on pystruct.
Cheers,
Also, it's not that GridSearch is sensitive in itself, but remember
you're doing LeaveOneOut, so for every grid point you are actually
doing `n_samples` calls to clf.fit.
Maybe one of these calls is significantly slower than others due to scaling.
On Wed, Jul 3, 2013 at 10:42 PM, Lars Buitinck
Why would autoencoders be naturally batch? I think historically one of
their early uses was for Online PCA, but I may be wrong.
Vlad
On Wed, Jun 26, 2013 at 11:51 PM, Issam issamo...@gmail.com wrote:
Hi @Olivier, you are absolutely right, scipy.optimize.fmin_l_bfgs_b
would not be suitable for
Hey everybody,
Today is the official starting date for GSoC 2013, and I am very excited!
As those of you following could definitely see, we had a lot of very good
proposals, sadly, more than we could accept. We managed to get 2 slots from
the PSF, and the projects that were
accepted are:
-
But using OrderedDict or some other Bunch 2.0 is beside the point. Even if
we find some awesome way of storing the datasets while allowing the cool
oneliner, it will still mislead people into thinking they need to put their
data into that format. What we want is to make it super-obvious that
Now, how to do that?
We don't. I am tired of completely dumming down our code to make it
usable by people who don't understand what they do. All it does is give
us extra work in terms of support.
In this case, we can't really do any better than the way it is, the Bunch
is pretty clear.
1 - 100 of 233 matches
Mail list logo