using k-means++ often means that the initialization takes much longer than the actual algorithm (for large n_clusters in praticular).
I wouldn't do cross-validation, and I'd see if you can make it faster by using less words or less iterations and get comparable performance.

I moved the dataset into example as suggested.

Regarding the parameters, I decrease the number of words to 80. I also limited the maximum number of patches extracted to 20,000 per images (maybe we will classify grass instead of cow), which is about 1/4-5th of the possible total number of patches.

The results seem appropriate with a large decrease of the computation time, which start to be suitable.

Classification performed - the confusion matrix obtained is:
[[23  0  0]
 [ 1 19  0]
 [ 0  0 23]]
It took 85.6792821884 seconds.

Using 50 words will make drop a bit the accuracy but the time is 35 sec.

Classification performed - the confusion matrix obtained is:
[[19  4  0]
 [ 0 20  0]
 [ 0  0 23]]
It took 33.9749679565 seconds.

Also, I am taking advantage of my 4 available threads (i7 2620M). I don't know if it has to be taken into account.

@GaelVaroquaux @amueller

What further improvements/changes should I take care of?

Cheers,

@amueller @GaelVaroquaux

What else should I look at?

sorry for the slow reply.
Please fix python3 compatibility, rename the example to plot_bovw.py" (theplot_`` is the important part, I just thought "example" was a bit redundant in the example folder ;).

Also add a header to the example and a text explaining what's happening.
thanks!

Ok, I am up to date now with the requirements. I'll do that in the week-end and ping once done.

@amueller Shall I remove the timer?

@amueller @GaelVaroquaux Do you see any other additional improvements to bring?

Should tudarmstadt.py be moved into the sklearn.datasets module.

I decided to decrease the number of patches to get under 10 sec of processing.
I got that confusion matrix which I think is fine for an example:

[[21  2  0]
 [ 4 16  0]
 [ 0  0 23]]

We are confusing cars and motorbike but we get the cows right ;D

Can you check the pep8 errors please?

Can you check the pep8 errors please?

@amueller I have the following errors since the doc come first.

plot_bovw.py:25:1: E402 module level import not at top of file
plot_bovw.py:26:1: E402 module level import not at top of file
plot_bovw.py:28:1: E402 module level import not at top of file
plot_bovw.py:29:1: E402 module level import not at top of file
plot_bovw.py:30:1: E402 module level import not at top of file
plot_bovw.py:31:1: E402 module level import not at top of file
plot_bovw.py:32:1: E402 module level import not at top of file
plot_bovw.py:33:1: E402 module level import not at top of file
plot_bovw.py:34:1: E402 module level import not at top of file
plot_bovw.py:36:1: E402 module level import not at top of file

@raghavrv I moved it. But I will recall what @amueller mentioned earlier

maybe I'll just add this to the example, and not put it in the datasets folder. This is not really a widely used dataset. and once it is in the public api, it is hard to get rid of.

That's why we kept it originally in the example folder. So let me know what is the best.

That's why we kept it originally in the example folder. So let me know what is the best.

Argh. My bad! I did not notice that comment... Could you very kindly revert please?

It has been hidden since that I previously addressed it. I remind it when it pop up after reversing.

@amueller ping

That's why we kept it originally in the example folder. So let me know what is the best.

Hmmm but then it kind of defeats the implicit contract of an example which is being able to copy and paste it in an IPython session and have it running without any hitch.

Not sure what is the best way of doing this (the dataset is 42MB so too big to be included in the scikit-learn repo). Here are a few alternatives I can think of:

• put the fetcher code in the example
• keep it like this but explain in the example docstring that you need to copy locally the fetcher file (potentially linking to its example HTML URL)
• create a sklearn.example_datasets for datasets that are only used in the examples

Not convinced by any of these alternatives, so better suggestions welcome!

Another alternative: you download the fetcher file tudarmstadt.py from the example gallery "Download Python file" URL and exec in inside plot_bovw.py. Keeps the example simple, maybe a little bit too much magic ...

I think that it needs either to be added inside the example or moved to the datasets module.

If we move it to the datasets module, do you foresee other examples built upon it?

I think that there is room for additional examples using this dataset and core methods from sklearn.
I would probably focus on examples proposed in the PASCAL challenge.

I don't think we'll have other examples use visual datasets (either this or PASCAL VOC) because we don't have the tools to extract the right features - and because everything runs waaaay to long for a scikit-learn example.

So therefore I'd rather not have it in the datasets folder - it's not very relevant to current computer vision, and it's not really helpful to show anything else in scikit-learn. But the example file is already quite long...

I don't think we'll have other examples use visual datasets (either this or PASCAL VOC) because we don't have the tools to extract the right features - and because everything runs waaaay to long for a scikit-learn example.

Fair enough. With this example, the only interesting point to me is the parallel between the Bag of Words and Bag of Visual Words.

@amueller are the last changes fine with you?

@amueller @GaelVaroquaux is this PR still of interest for merging or it should be closed?

I am closing that PR. It seems that scikit-learn has already plenty of examples.