Re: [Scikit-learn-general] partial_fit behavior for SGDClassifier

2013-10-15 Thread Gael Varoquaux
On Wed, Oct 16, 2013 at 01:40:09AM +0200, Jaques Grobler wrote: > I'm not 100% on this and would appreciate further input here, as I'm half > assleep right now! :D You are pretty much right: partial_fit does not try to iterate the model to convergence, because it expects to be called many times, p

Re: [Scikit-learn-general] partial_fit behavior for SGDClassifier

2013-10-15 Thread Jaques Grobler
Hi As far as I understand it, this seems expected to me. Basically, partial-fit does online-training on the data.. treating it's 'X' as a subset of the total data - so it expects to be working with chunks. when the second partial_fit call is made, it's not like a normal 'fit' that refits the classi

[Scikit-learn-general] partial_fit behavior for SGDClassifier

2013-10-15 Thread Akhil Shah
Hi, I'm using scikit-learn v0.14.1 and was wondering why classification results improve just because partial_fit was called a second time on the same training data? Note in the example below i'm using train/test overlap for debugging purposes. Other behavior I don't understand is if train/test sa

Re: [Scikit-learn-general] choice of regularization parameter grid for elastic net

2013-10-15 Thread Olivier Grisel
2013/10/15 Olivier Grisel : > 2013/10/15 Alexandre Gramfort : >>> I did find the part in coordinate_descent.py where alpha_max is chosen, but >>> I don't fully understand the reasoning behind it: >>> >>> alpha_max = np.abs(Xy).max() / (n_samples * l1_ratio) >> >> it can be derived from the KKT opti

Re: [Scikit-learn-general] choice of regularization parameter grid for elastic net

2013-10-15 Thread Olivier Grisel
2013/10/15 Alexandre Gramfort : >> I did find the part in coordinate_descent.py where alpha_max is chosen, but >> I don't fully understand the reasoning behind it: >> >> alpha_max = np.abs(Xy).max() / (n_samples * l1_ratio) > > it can be derived from the KKT optimality conditions of the Lasso probl

Re: [Scikit-learn-general] choice of regularization parameter grid for elastic net

2013-10-15 Thread Alexandre Gramfort
> I did find the part in coordinate_descent.py where alpha_max is chosen, but > I don't fully understand the reasoning behind it: > > alpha_max = np.abs(Xy).max() / (n_samples * l1_ratio) it can be derived from the KKT optimality conditions of the Lasso problem. A ---

Re: [Scikit-learn-general] Feature Filtering

2013-10-15 Thread Fred Mailhot
Use the same DictVectorizer that you called fit_transform() on with the training data, but just call transform() for the test data... dv = DictVectorizer() train_feats = dv.fit_transform(train_feature_dict) test_feats = dv.transform(test_feature_dict) On 15 October 2013 03:52, Lars Buitinck w

Re: [Scikit-learn-general] Feature Filtering

2013-10-15 Thread Lars Buitinck
2013/10/14 Osman Baskaya : > 2- In contrast to #1, I would like to automatically give 0 to those features > in training data that are not observed in the test data. > > If I have a feature that is not observed in the test set, I got error > because I am using the DictVectorizer so that it does not

[Scikit-learn-general] Feature Filtering

2013-10-15 Thread Osman Baskaya
I have two questions: 1- I would like to filter features in the test set that aren't observed in the training set. Is there a way to do it concise way? 2- In contrast to #1, I would like to automatically give 0 to those features in training data that are not observed in the test data. If I have

Re: [Scikit-learn-general] HMM with von Mises Emmissions

2013-10-15 Thread Robert McGibbon
Furthermore, I'm not sure that the API in seqlearn is the right fit for GaussianHMM, GMMHMM, and other models with continuous emission distributions. The sklearn HMM API is really geared towards unsupervised tasks. -Robert On Oct 15, 2013, at 12:57 AM, Lars Buitinck wrote: > 2013/10/15 Gael V

Re: [Scikit-learn-general] HMM with von Mises Emmissions

2013-10-15 Thread Robert McGibbon
I will take a look at t some of these tomorrow. -Robert On Oct 15, 2013, at 12:24 AM, Olivier Grisel wrote: > 2013/10/15 Fred Mailhot : >> On 14 October 2013 20:48, Robert McGibbon wrote: >> [...] >>> >>> >>> p.s. core devs: pretty please don't remove the HMM code from the scikit :) >> >> >

Re: [Scikit-learn-general] HMM with von Mises Emmissions

2013-10-15 Thread Lars Buitinck
2013/10/15 Gael Varoquaux : > What is important is that the functionality get's adopted in another > package. I don't know the exact scope of seqlearn, but it seems to me > that it might be a right home for HMMs. Maybe Lars can comment. I have little time to maintain seqlearn, but a PR is welcome.

Re: [Scikit-learn-general] HMM with von Mises Emmissions

2013-10-15 Thread Olivier Grisel
2013/10/15 Fred Mailhot : > On 14 October 2013 20:48, Robert McGibbon wrote: > [...] >> >> >> p.s. core devs: pretty please don't remove the HMM code from the scikit :) > > > +1E6 Would be great to have volunteers tackling HMM related issues then: https://github.com/scikit-learn/scikit-learn/se

Re: [Scikit-learn-general] HMM with von Mises Emmissions

2013-10-15 Thread Gael Varoquaux
On Mon, Oct 14, 2013 at 09:30:30PM -0700, Fred Mailhot wrote: >On 14 October 2013 20:48, Robert McGibbon <[1]rmcgi...@gmail.com> wrote: > [...] > p.s. core devs: pretty please don't remove the HMM code from the scikit > :) >+1E6 I think that it will happen. It's just a ques

Re: [Scikit-learn-general] recommendation systems

2013-10-15 Thread Mathieu Blondel
Adding recommendation systems also requires implementing new evaluation metrics and most likely new grid search tools. Mathieu On Mon, Oct 14, 2013 at 7:05 PM, Olivier Grisel wrote: > Actually the mrec implementation is not the original SLIM algorithm > but a variant demonstrated by the lib aut