[MRG] Export notebook to gallery #180
Conversation
|
I assume that in this PR you will also be fixing #141 |
|
Glad to see this going into documentation soon 💃 |
|
And this fixes #153 |
|
Regarding the dataset, should I go for the faces dataset for instance ? (Probably the supervised version since we use |
|
I just ran another |
Yes. Digits could also be an option. Note that you will have to use dimensionality reduction (say t-SNE) to visualize things in 2D. Otherwise we could work with a 2D or 3D dataset but there is not so many nice things to show in such cases |
Allright, I tried to launch the example with digits, but there is a conditionning problem with SDML, so I think we'll need to wait for #162 to be merged to finish this one |
|
Here is a first result of plotting with the faces dataset (with the digits dataset points are already well separated out with tsne on the raw dataset): tsne from the raw dataset: tsne after lmnn: tsne after sdml: tsne after lsml: tsne after nca: tsne after lfda: RCA failed (I still need to understand why) The improvement is not very clear, except maybe for lmnn, but for lfda for instance it's worse... Though it could work with appropriate tuning. |
|
I tried with the balance dataset and reducing the dimension gives a bit weird results (like I think balance kind of defines a grid of regularly spaced points, which seems a bit like an artificial dataset) (but anyway running the default version of the algos didn't work so much either) |
|
An option would be to add some noise dimensions (this is a bit artificial but will definitely work) Otherwise you have to look for harder datasets - if they t-SNE visualization is already very good on the original representation, it is hard to obtain clear visual improvements |
|
That's right, it works really well indeed with noise: Here are the images with 5 columns of gaussian noise N(0, 10) added to make_blobs dataset (the legend with sepal etc is wrong, but the image is good, it's the results after applying t-sne): tsne after LMNN: tsne after ITML: tsne after SDML: tsne after LSML: tsne after NCA: tsne after LFDA: tsne after RCA: So maybe we can use this with a quick intro paragraph saying "let's say we have a noisy dataset: let's take iris and add some columns of noise etc.." ? Otherwise do you know any dataset with a lot of noise already inside it ? Also, LSLM and SDML work not so good, so I guess as soon as we settle on the dataset I'll try to fine tune them and if it still doesn't work maybe we can remove them ? |
|
In the end I used iris with 5 columns of noise added, in the last commit is a version that I think is mergeable for the example, let me know what you think |
wdevazelhes
changed the title
|
Note that there is still |
|
I haven't read all the text yet, but I agree this is a good approach and a reasonable example to use. |
There was a problem hiding this comment.
I think it looks pretty good. As discussed it would be great to use make_classification from sklearn. This allows to natively have noisy dimensions, but also to have several clusters per class (which could be nice to illustrate the difference between methods based on local constraints like LMNN and those based on global ones like ITML), and many other options (class separation, label noise, etc).
|
I just pushed an example that uses @bellet I also wrote something about the fact that some algorithms don't try to cluster all points from the same class in one same cluster, while others implicitly do that, which can be seen quite clearly from these examples (since every class has a distribution with 2 clusters: If I understood correctly, the algorithm that don't cluster similar points in a unique cluster are |
|
If these changes are fine I think we are good to merge |
|
I just had a quick look and it looks good! I have a few nitpick suggestions on the wording and presentation, will write a review for that asap |
examples/plot_metric_examples.py
Outdated
| visualisation which can help understand which algorithm might be best | ||
| suited for you. | ||
|
|
||
| Of course, depending on the data set and the constraints your results |
examples/plot_metric_examples.py
Outdated
| ~~~~~~~~~~~~~~~~~~~~~~ | ||
|
|
||
| This is a small walkthrough which illustrates all the Metric Learning | ||
| algorithms implemented in metric-learn, and also does a quick |
There was a problem hiding this comment.
maybe rather "with some visualizations to provide intuitions into what they are designed to achieve".
examples/plot_metric_examples.py
Outdated
| ~~~~~~~~~~~~~~~~~~~~~~ | ||
|
|
||
| This is a small walkthrough which illustrates all the Metric Learning | ||
| algorithms implemented in metric-learn, and also does a quick |
There was a problem hiding this comment.
also I would mention that this is done on synthetic data
examples/plot_metric_examples.py
Outdated
| """ | ||
|
|
||
| # License: BSD 3 clause | ||
| # Authors: Bhargav Srinivasa Desikan <bhargavvader@gmail.com> |
There was a problem hiding this comment.
you should add yourself as an author
examples/plot_metric_examples.py
Outdated
| # Loading our data-set and setting up plotting | ||
| # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ | ||
| # | ||
| # We will be using the IRIS data-set to illustrate the plotting. You can |
There was a problem hiding this comment.
this is not iris anymore. update and link to make_classification documentation
There was a problem hiding this comment.
Thanks, I forgot to update this indeed
examples/plot_metric_examples.py
Outdated
| Algorithms walkthrough | ||
| ~~~~~~~~~~~~~~~~~~~~~~ | ||
|
|
||
| This is a small walkthrough which illustrates all the Metric Learning |
There was a problem hiding this comment.
Agreed, I will add MMC, but I don't know for MLKR, I think I'll just talk about it when describing NCA ? Since the cost function is very similar except that MLKR uses a soft-nearest neighbors for regression as I understood
There was a problem hiding this comment.
true, MLKR is for regression. it is a good idea to mention it there as proposed
maybe change "all the metric learning algorithms" to "most metric learning algorithms"
There was a problem hiding this comment.
In the end, I added most metric learning algorithms because we don't talk about Covariance, (we could maybe ?) And I added MLKR with make_regression task because I found the results were pretty cool too! Even if it breaks a bit the outline because it's just between supervised algorithms and the "constraints" section... Tell me what you think ?
examples/plot_metric_examples.py
Outdated
| # | ||
| # Implements an efficient sparse metric learning algorithm in high | ||
| # dimensional space via an :math:`l_1`-penalised log-determinant | ||
| # regularization. Compare to the most existing distance metric learning |
examples/plot_metric_examples.py
Outdated
| # Neighborhood Components Analysis | ||
| # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ | ||
| # | ||
| # NCA is an extrememly popular metric-learning algorithm, and one of the |
There was a problem hiding this comment.
extremely
maybe remove last part of the sentence
There was a problem hiding this comment.
Thanks, done for extremely
For the last part, is it because it can make people that don't know the algorithm think that if it's one of the first few it might be not cutting edge/outdated ?
Done
There was a problem hiding this comment.
just that it is maybe not such a relevant thing to mention here. you could also argue this for LMNN, and others like MMC
There was a problem hiding this comment.
That's right, they are also one of the first ones indeed
examples/plot_metric_examples.py
Outdated
|
|
||
|
|
||
| ###################################################################### | ||
| # Manual Constraints |
There was a problem hiding this comment.
We do not say explicitly what is meant by "constraint" before so I think this part is quite confusing
Also I think we want to insist on the fact that many metric learning algorithms only need the weak supervision given by constraints, not labels. In many applications, this is easier to obtain.
There was a problem hiding this comment.
I agree, I'll reformulate this paragraph, let me know what you think
There was a problem hiding this comment.
I think we can also delete the num_constraints=200 in all *_Supervised algorithms what do you think ?
There was a problem hiding this comment.
i guess so, unless it increases the computation time too much
There was a problem hiding this comment.
That's right, I tried and it tooked almost the same time, so I removed them
examples/plot_metric_examples.py
Outdated
| # it's worth one's while to poke around in the constraints.py file to see | ||
| # how exactly this is going on. | ||
| # | ||
| # This brings us to the end of this tutorial! Have fun Metric Learning :) |
There was a problem hiding this comment.
maybe also add that metric-learn is compatible with sklearn and so we can easily do model selection, cross validation, scoring etc and refer to the doc for more details.
|
@bellet Thanks for your review, I addressed your comments |
examples/plot_metric_examples.py
Outdated
| # distance in the input space, in which the contribution of the noisy | ||
| # features is high. So even if points from the same class are close to | ||
| # each other in some subspace of the input space, this is not the case in the | ||
| # total input space. |
There was a problem hiding this comment.
maybe "this is not the case when considering all dimensions of the input space"
examples/plot_metric_examples.py
Outdated
| # good literature review of Metric Learning. | ||
| # | ||
| # We will briefly explain the metric-learning algorithms implemented by | ||
| # metric-learn, before providing some examples for it's usage, and also |
examples/plot_metric_examples.py
Outdated
| # | ||
| # Basically, we learn this distance: | ||
| # :math:`D(x,y)=\sqrt{(x-y)\,M^{-1}(x-y)}`. And we learn this distance by | ||
| # learning a Matrix :math:`M`, based on certain constraints. |
There was a problem hiding this comment.
- there is repetition here "we learn this distance"
- add very quick explanation of what is meant by "constraint" (can be inspired from What is metric learning? page). For instance something like, "we learn the parameters :math:
Mof this distance to satisfy certain constraints on the distance between points, for example requiring that points of the same class are close together and points of different class are far away."
examples/plot_metric_examples.py
Outdated
| # | ||
| # We will briefly explain the metric-learning algorithms implemented by | ||
| # metric-learn, before providing some examples for it's usage, and also | ||
| # discuss how to go about doing manual constraints. |
There was a problem hiding this comment.
remove manual constraint and replace by something like "discuss how to perform metric learning with weaker supervision than class labels"
examples/plot_metric_examples.py
Outdated
| plot_tsne(X_rca, Y) | ||
|
|
||
| ###################################################################### | ||
| # Metric Learning for Kernel Regression |
There was a problem hiding this comment.
Regression example: Metric Learning for Kernel Regression
examples/plot_metric_examples.py
Outdated
| # going to go ahead and assume that two points labelled the same will be | ||
| # closer than two points in different labels. | ||
| # | ||
| # Do keep in mind that we are doing this method because we know the labels |
There was a problem hiding this comment.
move this above, right after the sentence saying that we are going to create constraints from the labels
examples/plot_metric_examples.py
Outdated
|
|
||
|
|
||
| ###################################################################### | ||
| # Using our constraints, let's now train ITML again. We should keep in |
There was a problem hiding this comment.
not sure the last sentence is needed (already said that before)
examples/plot_metric_examples.py
Outdated
| ###################################################################### | ||
| # And that's the result of ITML after being trained on our manual | ||
| # constraints! A bit different from our old result but not too different. | ||
| # We can also notice that it might be better to rely on the randomised |
There was a problem hiding this comment.
this last sentence is not very clear. i would remove it
examples/plot_metric_examples.py
Outdated
| # also compatible with scikit-learn, since their input dataset format described | ||
| # above allows to be sliced along the first dimension when doing | ||
| # cross-validations (see also this :ref:`section <sklearn_compat_ws>`). See | ||
| # also some :ref:`use cases <use_cases>` where you could use scikit-learn |
There was a problem hiding this comment.
where you could combine metric learning with scikit-learn estimators
examples/plot_metric_examples.py
Outdated
| # pipeline or cross-validation procedure. And weakly-supervised estimators are | ||
| # also compatible with scikit-learn, since their input dataset format described | ||
| # above allows to be sliced along the first dimension when doing | ||
| # cross-validations (see also this :ref:`section <sklearn_compat_ws>`). See |
|
Thanks for the review @bellet, I addressed all your comments |
|
Thanks! I have pushed a few small updates myself to avoid another reviewing cycle. I think we can merge once CI passes |
Fixes #141 #153
Hi, I've just converted @bhargavvader 's notebook from #27 into a sphinx-gallery file (with this snippet: https://gist.github.com/chsasank/7218ca16f8d022e02a9c0deb94a310fe). This way, it will appear nicely in the documentation, and can also allow us to check if every algorithms work fine.
There are a few things to change to make the PR mergeable (to compile the doc, you need sphinx-gallery):