Mailman 3 November 2016 - scikit-learn

Github project management tools
by Joel Nothman Dec. 3, 2016

Dec. 3, 2016

One of the biggest issues with scikit-learn as a project is managing its backlog of issues; another is release scheduling. Some of this cannot be fixed as long as our model of voluntary contribution (with a couple of important exceptions) does not change. However, it may be worth considering the new project management features in Github. At the moment we have the following management: * labels corresponding to type (bug, enhancement, new feat, question), scope (API, Build/CI, ?Large Scale, Documentation), difficulty (easy, moderate), status/scheduling (needs contributor, needs review, sprint). * PR status management with title prefixes [WIP], [MRG], [MRG+1], [MRG+2] Firstly, we might benefit from prefixing labels by category, i.e. difficulty:easy so that complementary labels appear together. In truth, PRs have roughly these statuses: * WIP (not ready for review) * waiting for review * waiting for changes (with or without one of the following) * in dispute (i.e. fundamental doubts about the PR) * the above together with 1 or 2 "official" approvals * ready for merge (pending minor changes such as what's new documentation) New github features: * reviews with "approved" or "request changes". A list of approvers can be found in the merge/CI panel. We could replace the MRG+1 annotation with this and use it to track disputation too. I'm not sure how it works with changes that are added after approval. I think it would have avoided one improper merge by me... One downside is that there does not yet seem to be a way to search for PRs with a specified level of approval (while searching for "MRG+1" sort-of works). * Milestone prioritising: issues in a milestone, such as https://github.com/scikit-learn/scikit-learn/milestone/21, can be ranked with drag-and-drop. I think this could help with release scheduling as it would allow us to identify the top priorities for a release and see when enough of them are completed. * The Kanban-style workflow management of the new Projects tool https://github.com/scikit-learn/scikit-learn/projects is another way of managing status and, I think, priority, for a small set of related issues. This might be an alternative way of managing milestone scope, or of working towards big changes like the one just completed for model selection; like proposed expansions to get_feature_names expansion; like estimator tags; making utilities public/private... So with the goal of making it easier to track where attention is most needed, and when to move to release: What's worth trying?

13 37

Bayesian Gaussian Mixture
by Tommaso Costanzo Nov. 30, 2016

Nov. 30, 2016

Hi, I am facing some problem with the "BayesianGaussianMixture" function, but I do not know if it is because of my poor knowledge on this type of statistics or if it is something related to the algorithm. I have set of data of around 1000 to 4000 observation (every feature is a spectrum of around 200 point) so in the end I have n_samples = ~1000 and n_features = ~20. The good things is that I am getting the same results of KMeans however the "predict_proba" has value only of 0 or 1. I have wrote a small function to simulate my problem with random data that is reported below. The first 1/2 of the array has the point with a positive slope while the second 1/2 has a negative slope, so the cross in the middle. What I have seen is that for a small number of features I obtain good probability, but if the number of features increases (say 50) than the probability become only 0 or 1. Can someone help me in interpret this result? Here is the code I wrote with the generated random number, I'll generally run it with ncomponent=2 and nfeatures=5 or 10 or 50 or 100. I am not sure if it will work in every case is not very highly tested. I have also attached as a file! ########################################################################## import numpy as np from sklearn.mixture import GaussianMixture, BayesianGaussianMixture import matplotlib.pyplot as plt def test_bgm(ncomponent, nfeatures): temp = np.random.randn(500,nfeatures) temp = temp + np.arange(-1,1, 2.0/nfeatures) temp1 = np.random.randn(400,nfeatures) temp1 = temp1 + np.arange(1,-1, (-2.0/nfeatures)) X = np.vstack((temp, temp1)) bgm = BayesianGaussianMixture(ncomponent,degrees_of_freedom_prior=nfeatures*2).fit(X) bgm_proba = bgm.predict_proba(X) bgm_labels = bgm.predict(X) plt.figure(-1) plt.imshow(bgm_labels.reshape(30,-1), origin='lower', interpolatio='none') plt.colorbar() for i in np.arange(0,ncomponent): plt.figure(i) plt.imshow(bgm_proba[:,i].reshape(30,-1), origin='lower', interpolatio='none') plt.colorbar() plt.show() ############################################################################## Thank you in advance Tommaso -- Please do NOT send Microsoft Office Attachments: http://www.gnu.org/philosophy/no-word-attachments.html

3 5

Problem with nested cross-validation example?
by Daniel Homola Nov. 30, 2016

Nov. 30, 2016

Dear all, I was wondering if the following example code is valid: http://scikit-learn.org/stable/auto_examples/model_ selection/plot_nested_cross_validation_iris.html My understanding is, that the point of nested cross-validation is to prevent any data leakage from the inner grid-search/param optimization CV loop into the outer model evaluation CV loop. This could be achieved if the outer CV loop's test data is completely separated from the inner loop's CV, as shown here: https://mlr-org.github.io/mlr-tutorial/release/html/img/ nested_resampling.png The code in the above example however doesn't seem to achieve this in any way. Am I missing something here? Thanks a lot, dh

6 13

Bugs in Tree.py
by ALVARENGA GAMERO ALFONSO ABRAHAM Nov. 29, 2016

Nov. 29, 2016

sklearn/tree/tree.pyWith the new 0.18 version, it is possible to add percentages values for "min_samples_split"#.. versionchanged:: 0.18#Added float values for percentages.How ever, a value of 1 will make the program to issue an ValueError (lines 195-199), since 1 is an Integer and does not hold the condition of being bigge ror equal than 2. It is quite easy to solve by hand (if not 2 <= self.min_samples_split and self.min_samples_split != 1: in line 196), but I'm pretty sure there has to be a clever way to solve it. I might go back to that later, as there might be more bugs as this one with the new options in version 0.18.Thank you!

3 2

How to not recalculate transformer in a Pipeline?
by Anton Suchaneck Nov. 29, 2016

Nov. 29, 2016

Hello! I use a 2-step Pipeline with an expensive transformer and a classification afterwards. On this I do GridSearchCV of the classifcation parameters. Now, theoretically GridSearchCV could know that I'm not touching any parameters of the transformer and avoid re-doing work by keeping the transformed X, right?! Currently, GridSearchCV will do a clean re-run of all Pipeline steps? Can you recommend the easiest way for me to use GridSearchCV+Pipeline while avoiding recomputation of all transformer steps whose parameters are not in the GridSearch? I realize this may be tricky, but any pointers to realize this most conveniently and compatible with sklearn would be highly appreciated! (The scoring has to be done on the initial data, so I cannot just manually transform beforehand.) Regards, Anton PS: If that all makes sense, is that a useful feature to include in sklearn?

4 10

Specifying exceptions to ParameterGrid
by Jaidev Deshpande Nov. 26, 2016

Nov. 26, 2016

Hi, Sometimes when using GridSearchCV, I realize that in the grid there are certain combinations of hyperparameters that are either incompatible or redundant. For example, when using an MLP, if I specify the following grid: grid = {'solver': ['sgd', 'adam'], 'learning_rate': ['constant', 'invscaling', 'adaptive']} then it yields the following ParameterGrid: [{'learning_rate': 'constant', 'solver': 'sgd'}, {'learning_rate': 'constant', 'solver': 'adam'}, {'learning_rate': 'invscaling', 'solver': 'sgd'}, {'learning_rate': 'invscaling', 'solver': 'adam'}, {'learning_rate': 'adaptive', 'solver': 'sgd'}, {'learning_rate': 'adaptive', 'solver': 'adam'}] Now, three of these are redundant, since learning_rate is used only for the sgd solver. Ideally I'd like to specify these cases upfront, and for that I have a simple hack ( https://github.com/jaidevd/jarvis/blob/master/jarvis/cross_validation.py#L38). Using that yields a ParameterGrid as follows: [{'learning_rate': 'constant', 'solver': 'adam'}, {'learning_rate': 'invscaling', 'solver': 'adam'}, {'learning_rate': 'adaptive', 'solver': 'adam'}] which is then simply removed from the original ParameterGrid. I wonder if there's a simpler way of doing this. Would it help if we had an additional parameter (something like "grid_exceptions") in GridSearchCV, which would remove these dicts from the list of parameters? Thanks

4 9

question about using sklearn.neural_network.MLPClassifier？
by linjia＠ruijie.com.cn Nov. 25, 2016

Nov. 25, 2016

Hi everyone I try to use sklearn.neural_network.MLPClassifier to test the XOR operation, but I found the result is not satisfied. The following is code, can you tell me if I use the lib incorrectly? from sklearn.neural_network import MLPClassifier X = [[0, 0], [0, 1], [1, 0], [1, 1]] y = [0, 1, 1, 0] clf = MLPClassifier(solver='adam', activation='logistic', alpha=1e-3, hidden_layer_sizes=(2,), max_iter=1000) clf.fit(X, y) res = clf.predict([[0, 0], [0, 1], [1, 0], [1, 1]]) print(res) #result is [0 0 0 0], score is 0.5

3 8

Development workflow proposal: merge master instead of rebasing
by Andreas Mueller Nov. 20, 2016

Nov. 20, 2016

Hi all. I think we should change our development practices for resolving merge-conflicts from rebasing to merging. The "squash and merge" button of github gets rid of any merge commits and results in a clean history in any case. The benefit of merging instead of rebasing is that github is able to track comments much better if you don't force-push. In particular the links in notification emails might work better when not doing force-pushes. I'm not entirely sure how the mechanism works, but I think it's worth giving it a go. When merging master it's also harder to screw up an PR entirely (I think) which would make it easier to people new to git. Wdyt? Andy

4 4

suggested classification algorithm
by Thomas Evangelidis Nov. 17, 2016

Nov. 17, 2016

Greetings, I want to design a program that can deal with classification problems of the same type, where the number of positive observations is small but the number of negative much larger. Speaking with numbers, the number of positive observations could range usually between 2 to 20 and the number of negative could be at least x30 times larger. The number of features could be between 2 and 20 too, but that could be reduced using feature selection and elimination algorithms. I 've read in the documentation that some algorithms like the SVM are still effective when the number of dimensions is greater than the number of samples, but I am not sure if they are suitable for my case. Moreover, according to this Figure, the Nearest Neighbors is the best and second is the RBF SVM: http://scikit-learn.org/stable/_images/sphx_glr _plot_classifier_comparison_001.png However, I assume that Nearest Neighbors would not be effective in my case where the number of positive observations is very low. For these reasons I would like to know your expert opinion about which classification algorithm should I try first. thanks in advance Thomas -- ====================================================================== Thomas Evangelidis Research Specialist CEITEC - Central European Institute of Technology Masaryk University Kamenice 5/A35/1S081, 62500 Brno, Czech Republic email: tevang(a)pharm.uoa.gr tevang3(a)gmail.com website: https://sites.google.com/site/thomasevangelidishomepage/

5 7

Including figures from scikit-learn documentation in scientific publications
by avn＠mccme.ru Nov. 16, 2016

Nov. 16, 2016

Hello, I'm writing a paper meant for submission to TPAMI and would like to include that wonderful figure of clustering algorithms comparison found in scikit-learn documentation (http://scikit-learn.org/stable/_images/sphx_glr_plot_cluster_comparison_001…). So, my question is: can I include the figure PNG file in my paper directly (with proper reference, of course) or should I only provide a reference to this figure? With best regards, -- Valery

3 4