Domain adaptation Classes #22

Slasnista · 2017-07-28T13:47:57Z

first proposal of DA class structure
BaseEstimator: OTDA wrapper (does not work as a stand-alone but implements the methods common to any OTDA algorithm)
SinkhornTransport: implements Sinkhorn algorithm for OTDA
try doc strings compliant with numpy requirements

rflamary · 2017-07-28T14:21:46Z

Thank you @Slasnista , I will have a look this week end or monday.

Maybe creating a compliant BaseEstimator Class as @agramfort proposed is a nice way to avaoid sklearn dependency.

rflamary · 2017-07-31T07:42:25Z

Hello, I left a few comments in response to @ngayraud.

@Slasnista I think this is a good start but we need to agree on a typical use case. below how I see the class working:

tr=EMDTransport(out_of_sample_map='ferradans',weight_est='uniform')

tf.fit(Xs,Xt) # learn G matrix and store it, also store data

Xs_trans=tr.transform(Xs) # barycentric mapping
Xs_trans=tr.transform() # barycentric mapping with stored data
Xs2_trans=tr.transform(Xs2) # barycentric mapping + Ferradans since out of sample

Slasnista · 2017-07-31T08:38:39Z

@rflamary thanks for the comments. I took them into account and I will update my pr.

For the use cases I agree with you on most of them. Yet even it's a bit redundant, it looks weird to not pass data to the transform method.

So if we opt for this solution, we need to write a detailed example to explain the differences between the different version of transform. I will look at your examples to update them accordingly to the classes structure.

rflamary · 2017-07-31T08:43:16Z

@Slasnista OK, as a second though I agree with you it's weird with no params. Let's not do it.

Still I think it is very important to handle automatically the known Xs and new our of samples Xs2.

Rémi

Slasnista · 2017-07-31T09:01:08Z

changes

as @agramfort suggested, I removed dependency on sklearn's BaseEstimator
removed also dependency on sklearn's pairwise_distance as @ngayraud proposed
added in __init__a parameter density_estimation which takes a function as input. (default: uniform)
stored Xs and Xt used for coupling estimation as attributes.
transform and inverse_transform modified: they check if their respective input has already been seen: if yes applies barycentric mapping, if no they will apply out of samples mapping (not implemented yet)

rflamary

This is nice, just a few more details and we have a workable class.

Think about adding somme test to test_da.py

rflamary · 2017-07-31T09:03:45Z

ot/da.py

+##############################################################################
+# adapted from scikit-learn
+
+import warnings


feel free to put it at the top of the file

rflamary · 2017-07-31T09:06:46Z

ot/da.py

+    # ``if type(self).__module__.startswith('sklearn.')``.
+
+
+def distribution_estimation_uniform(X):


note that there exist the function ot.utils.unif that does exactly the same thing

maybe a simple

from .utils import unif distribution_estimation_uniform=unif

would do the trick

a no it does not wrok but you still can use unif at the return line

rflamary · 2017-07-31T09:09:24Z

ot/da.py

+
+        # store arrays of samples
+        self.Xs = Xs
+        self.Xt = Xt


Remove sinkhorn from the BAseEstimator and put it in the SinkhornTransport please.

Keep all teh rest so that we can do super as still have all the storing and density estimation.

ngayraud · 2017-07-31T09:59:06Z

ot/da.py

@@ -1114,7 +1255,10 @@ class SinkhornTransport(BaseTransport):

    def __init__(self, reg_e=1., mode="unsupervised", max_iter=1000,
                 tol=10e-9, verbose=False, log=False, mapping="barycentric",
-                 metric="sqeuclidean", distribution="uniform"):
+                 metric="sqeuclidean",
+                 distribution_estimation=distribution_estimation_uniform,


Ok good. It is much simpler and allows the user to move around freely with the distribution estimation.

ngayraud · 2017-07-31T10:01:42Z

ot/da.py

        self.method = "sinkhorn"
+        self.out_of_sample_map = out_of_sample_map

    def fit(self, Xs=None, ys=None, Xt=None, yt=None):
        """Build a coupling matrix from source and target sets of samples


I see that in every fit and transform, all of the parameters are initialized to None. However, I think that at least Xs (and Xt for the inverse transform) should be required.

Slasnista · 2017-08-01T08:56:34Z

Hi,

I added a test function for SinkhornTransport()
I slightly changed SinkhornTransport() class. There are still some changes needed but we are converging.

Here some questions:

could we move the test folder in the ot foler ?
why are there so many things in the __init__ of ot folder ?
could we change the names of parameters and variables in da functions ? (I fear I, G, M... might be tricky for debugging in few months ?)
do you have a way to check the rendering of doc string ? I'd like to make new class docstrings compliant with your guidelines now, so that we won't have to spend extra sessions documenting code ;)

rflamary

OK thank you @Slasnista , it's indeed starting to converge and i see that test are not failing anymore !

before merging I would like to see at least the EMDTransport() implemented. Ican take care of the classes whth linear and kernel mapping estimation if you want.

I added a little comment and I will answer your questions:

Not in thus PR we cannot. This can be discussed in the future but this PR focus on a new and improved class for ot.
there are a lot of things imported because as for numpy I want all the base ot functions (emd,emd2,sinkhorn,sinkhorn2, dist) to be easily abailable with no need for importing submodule.
variable name change will not be done in this PR also because it is a full breaker for legacy classes ans all examples. It is open for discussion but note that ALL the ot function in the toolbox use this denomination and consistency is IMHO very important. Finally G/gamma/T are standard names for an OT matrix in the ot community and M/C are stanadr names for cost matrices.
on spyder editor, you can perform a ctrl-i to open the documentation of the current function (just click on its definition before). the documentation can be seen in raw form of with rendering.

rflamary · 2017-08-01T09:56:31Z

test/test_da.py

+np.random.seed(42)
+
+
+def test_sinkhorn_transport():


maybe test_sinkhorn_transport_class would be a better name

thank you for the reply, it's good for me. I will modify the name of the test function

ngayraud · 2017-08-01T10:43:25Z

Here is a bit of code you can add for a simple EMD transport. I hope I have it right.
Also from what I remember, if we set log = true, the methods return the log as well. So we need an "If log statement (see code below)

class EMDTransport(BaseTransport):
    """Domain Adapatation OT method based on Earth Mover's Distance
    Parameters
    ----------
    mode : string, optional (default="unsupervised")
        The DA mode. If "unsupervised" no target labels are taken into account
        to modify the cost matrix. If "semisupervised" the target labels
        are taken into account to set coefficients of the pairwise distance
        matrix to 0 for row and columns indices that correspond to source and
        target samples which share the same labels.
    mapping : string, optional (default="barycentric")
        The kind of mapping to apply to transport samples from a domain into
        another one.
        if "barycentric" only the samples used to estimate the coupling can
        be transported from a domain to another one.
    metric : string, optional (default="sqeuclidean")
        The ground metric for the Wasserstein problem
    distribution : string, optional (default="uniform")
        The kind of distribution estimation to employ
    verbose : int, optional (default=0)
        Controls the verbosity of the optimization algorithm
    log : int, optional (default=0)
        Controls the logs of the optimization algorithm
    Attributes
    ----------
    References
    ----------
    .. [1] N. Courty; R. Flamary; D. Tuia; A. Rakotomamonjy,
           "Optimal Transport for Domain Adaptation," in IEEE Transactions
           on Pattern Analysis and Machine Intelligence , vol.PP, no.99, pp.1-1
    """

    def __init__(self, mode="unsupervised", verbose=False, 
                 log=False, metric="sqeuclidean",
                 distribution_estimation=distribution_estimation_uniform,
                 out_of_sample_map='ferradans'):

        self.mode = mode
        self.verbose = verbose
        self.log = log
        self.metric = metric
        self.distribution_estimation = distribution_estimation
        self.out_of_sample_map = out_of_sample_map

    def fit(self, Xs, ys=None, Xt=None, yt=None):
        """Build a coupling matrix from source and target sets of samples
        (Xs, ys) and (Xt, yt)
        Parameters
        ----------
        Xs : array-like of shape = [n_source_samples, n_features]
            The training input samples.
        ys : array-like, shape = [n_source_samples]
            The class labels
        Xt : array-like of shape = [n_target_samples, n_features]
            The training input samples.
        yt : array-like, shape = [n_labeled_target_samples]
            The class labels
        Returns
        -------
        self : object
            Returns self.
        """

        self = super(EMDTransport, self).fit(Xs, ys, Xt, yt)

        # coupling estimation
        if self.log:
                    self.gamma_, self.log_ = emd(
                        a=self.mu_s, b=self.mu_t, M=self.Cost, 
                        verbose=self.verbose, log=self.log)
        else:
                    self.gamma_= emd(
                        a=self.mu_s, b=self.mu_t, M=self.Cost, 
                        verbose=self.verbose, log=self.log)

Slasnista · 2017-08-01T11:01:03Z

@ngayraud thank you for the code ! I adapt it and put it inside the pr ;)

just for if we can solve it like it

# coupling estimation
returned_ = emd(
    a=self.mu_s, b=self.mu_t, M=self.Cost, 
    verbose=self.verbose, log=self.log)
if self.log:
     self.coupling_, self.log = returned_
else:
    self.coupling_ = returned_

That does not change much, but we do not have the coupling estimation in a loop any more.

Btw, do all the other methods return log in addition to couplings ?

rflamary · 2017-08-01T11:14:30Z

Emd do not accept log or verbose. It's juste à frapper for c code. Le 1 août 2017 13:01, "Stanislas Chambon" <notifications@github.com> a écrit :

…

@ngayraud <https://github.com/ngayraud> thank you for the code ! I adapt it and put it inside the pr ;) just for if we can solve it like it # coupling estimation returned_ = emd( a=self.mu_s, b=self.mu_t, M=self.Cost, verbose=self.verbose, log=self.log) if self.log: self.coupling_, self.log = returned_ else: self.coupling_ = returned_ That does not change much, but we do not have the coupling estimation in a loop any more. Btw, do all the other methods return log in addition to couplings ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#22 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABUpecXzZWLWL2W5mIs-swJCzwG1Py68ks5sTwVvgaJpZM4Omn_M> .

Slasnista · 2017-08-01T11:15:49Z

@rflamary thx, the test gave me this error ;) I removed the log and verbose input to emd for the moment.

ngayraud · 2017-08-01T11:26:45Z

I think that all the sinkhorn based transports return a log! I was unsure about the emd one though.

…ng return self)

Slasnista · 2017-08-04T10:11:19Z

Hi,
I did some update to the PR:

add a test for fit_transform method and correction of a bug for this method (did not return self)
add 2 new classes with class regularization SinkhornLpl1Transport() and SinkhornL1l2Transport()
added few in lines in .fit() method of BaseTransport() class to support semi supervised domain adaptation. It implies that the labeled sources (resp. target) samples are in the first rows (resp. columns) + added a test to check that when semi supervised is selected but no labels (neither ys, yt, nor both) unsupervised DA is performed instead of semi-supervised

What do you think about these new modifications ?

rflamary · 2017-08-04T10:47:54Z

Hello @Slasnista,

Those are nice new features.

I just have a few questions/todos:

Why is the method necessary? I think the fact that the ys and yt are set to None is enough to know if semi supervised should be used. No need to have two signal that can contradict...
Also what is that about the supervised sample being the first ones? Couldn't we use ys/yt to define the supervised ones (0..N) and usnupervised (-1) ?
Semi supervised as in the PAMI paper does not work with O cost when the class is the same. It uses infinite (or very large) cost when the label is different. This is a versy different approach based on DO-NOT-LINK constratints and still provide OT inside classes which is no done when the loss in 0.
Finally I think the last piece before merging is coding a proper out of sample as done here:
https://github.com/Slasnista/POT/blob/0b005906f9d78adbf4d52d2ea9610eb3fde96a7c/ot/da.py#L712
Note that it's just a few lines but one should perform it on minibach in order to limit the memory requirement and hence it should Have a loop.
I know I am never happy but the interpolation (linear and kernel regression methods are also missing).
Once all the classes are done properly, feel free to add a warning to OTDA class that tell about the future deprecation and the new classes. I will add both of you to the contributor list BTW.

This will be a very nice merge thank you again for all the work .

Slasnista · 2017-08-04T11:28:33Z

@rflamary:

ok for me, this way it will not be too redundant.
ok, the way it's written works too independently of the position of the labeled samples so nothing to change
ok, but how do you set this infinite value ? 10**2, 10**3, something based on the max value of the cost matrix ? Do you have a predefined rule for doing this ?

I address the following points in a a second time ;)

rflamary · 2017-08-04T11:37:33Z

@Slasnista nice to see we agree. Point 3 is the tough one ;).

For sinkhorn and group lasso regularization we set it to infinity (np.inf) and under constraint qualification conditions (there exist a solution that is not infinite), the algorithm converges to a solution (with zerso mass on the do-not-link links).

For emd solver (and conjugate gradient that uses emd) I don't think the solver is going to appreciate infinity so I would use as you propose the max value of the cost matrix plus something maybe epsilon(max(Cost)). I think it should be a parameter anyway with the max/inf taken if set to None.

Take your time for the remaining I will be in vacation next week and probably won't fine time to do a proper final review anyway.

Slasnista · 2017-08-23T14:52:50Z

I completed all the modifications we talked about.

I think we should grasp this occasion to think about the algorithms' outputs. In my opinion:

functions which compute couplings and or mappings should always return logs_.
functions which compute couplings and mappings should always return coupling_, mapping_, bias_, logs_. I would prefer return a zero vector instead of having an output of varying size depending of the bias.

Point 2. is not a burning issue but 1. is one.

Right now, instantiating an object with logs=True results in an error. I propose to remove logs from __init__ and to let all .fit methods create an attribute .logs_ which will contain the logs. What do you think about it ?

rflamary · 2017-08-25T06:56:05Z

Hello @Slasnista,

Thank you for the new commits that handle all the things we discussed. I will do a proper code review but in the meantime you need to update the examples in the examples folder that use the old classes:

plot_OTDA_classes.py
plot_OTDA_color_images.py
plot_OTDA_mapping.py
plot_OTDA_mapping_color_images.py

We don't want the examples to be deprecated ;) Once this is done I will take care of updating the notebooks which is not done automatically yet.

About your questions, there should be a log parameter in init of the classes (note that it's log and not logs in all the toolbox as a parameter, log_ should be ok for storing the log ) and not the fit. Depending on this parameter the fit method should update the log_ parameter in the class. This parameter should exist all the time and be a an empty dictionary when log is set to False (maybe it should be set in the init of the Base class ?).

I understand that the varying size return in the algorithm is surprising but it leads to a very simple use of the functions while still allowing expert user to get more information if needed. Changing the behavior of the algorithm now will completely break the toolbox and again is not the object of this PR.
The reason for this is that computing the log can be time consuming since more things have to be returned (and the user should be able to choose of not to compute it).

ncourty · 2017-08-25T08:23:05Z

Hi all, entering the discussion after some holiday break :)
@rflamary, regarding the notebooks, since we decided to go full sphinx-gallery for the example, it may be the right tempo to consider re-writing them in the sphinx syntax ?

agramfort · 2017-08-25T08:30:56Z

" in the sphinx syntax" you mean in plain python :)

Slasnista · 2017-08-25T08:34:06Z

Hi @rflamary,

there si no error anymore when log=True
I created a .log_ that is either filled or instantiated as an empty dictionary depending on whether log is set to True
for methods without log parameter (emd, lpl1) I removed log from parameters in the __init__
I added some test to check that log_ is filled when log=True.

rflamary · 2017-08-25T11:43:08Z

Hello everyone,

@ncourty since you are back from holiday and full of energy maybe you could take care of the test file formating ;) ? As POT maintainer you should be able to push commits on the PR.

@Slasnista thank you, every thing looks great, I will do a proper code review shortly but I think we will be able to do a merge shortly.

Slasnista · 2017-08-25T13:18:29Z

Done with the examples according to the new OTDA objects !

@rflamary you're welcome, I now look forward to seeing it merged ;-)

rflamary

OK, just a few changes and a go from travis and I think we can merge and do a new POT version.

@Slasnista don't forget to add yourself and @ngayraud to the contributors of the toolbox in the readme file.

rflamary · 2017-08-28T07:24:00Z

ot/deprecation.py

@@ -0,0 +1,103 @@
+"""


Personally I think this short module is a bit overkill, maybe move the class and function to the utils.py module with the origin URL.

Ok I do that

rflamary · 2017-08-28T07:27:17Z

ot/da.py

@@ -10,21 +10,28 @@
 # License: MIT License

 import numpy as np
+import warnings


is it still necessary? I think everything is in deprecated now.

I'll check whether it's used in the script otherwise I remove it

Oh actually I need into the BaseEstimator class.

I will move the BaseEstimator classe into utils.py. That would be more relevant

rflamary · 2017-08-28T07:28:45Z

ot/da.py

@@ -706,12 +763,19 @@ def normalizeM(self, norm):
            self.M = np.log(1 + np.log(1 + self.M))


+@deprecated("The class OTDA_sinkhorn is deprecated in 0.3.1 and will be"


Love the decorator! Removed in 0.5 is short but probably a good thing since it will disappear from the examples shortly

rflamary · 2017-08-28T07:30:08Z

examples/da/plot_otda_classes.py

@@ -0,0 +1,142 @@
+# -*- coding: utf-8 -*-


You chose to move the da examples in a sub-folder do you think it is more clear or is it to be similar to sk-learn? did you remove the old ones, every trace of the old class should disappear from the examples?

I think sub-categories would clearer for examples and later on for the doc generation

I did not remove the old ones, shoud I move them in the new sub-folder ? @agramfort : do you have some advice to deprecate examples / documention jointly with some code ?

I think we shoud remove them from the examples since it's what people look at when trying out the toolbox.

Ok I do that so

yes remove deprecated code from all examples and documentation

rflamary · 2017-08-28T07:31:26Z

ot/da.py

+                      if p.name != 'self' and p.kind != p.VAR_KEYWORD]
+        for p in parameters:
+            if p.kind == p.VAR_POSITIONAL:
+                raise RuntimeError("scikit-learn estimators should always "


remove scikit-learn and say POT ;)

rflamary · 2017-08-28T07:36:03Z

ot/da.py

+# adapted from sklearn
+
+class BaseEstimator(object):
+    """Base class for all estimators in scikit-learn


add POt reference and a link to the original source in sk-learn

oh yes, i remove all the scikit refs like this one

rflamary · 2017-08-28T07:37:46Z

ot/da.py

+
+
+class BaseTransport(BaseEstimator):
+


Add documentation and list of necessary parameters in init for the class to run its functions
such as metric

I do not do that for BaseTransport since it has no __init__ and no real parameters but I will check the others to not forget parameters

what i'm saying is that you code here functions fit and transform that require some parameters/properties to be set before calling ()hence that should be defined in init. you should express here which properties should be set in order for the function to work.

this to allows new contributors to propose new classes with no headaches ;)

rflamary · 2017-08-28T07:40:55Z

ot/da.py

+
+class BaseTransport(BaseEstimator):
+
+    def fit(self, Xs=None, ys=None, Xt=None, yt=None):


Since we give all None values to the input parameters, we should handle those in case the user just call the function with nothing.

When a necessary parameter is missing we should print a warning and not do anything.

Good remark, I will handle the possible exceptions

rflamary · 2017-08-28T07:46:59Z

ot/da.py

+
+        return self
+
+    def fit_transform(self, Xs=None, ys=None, Xt=None, yt=None):


SAme here with the necessary parameters, maybe we should coce a test function with an error message that lists the necessary parameters i'm thinking as

def check_params(**kwargs): OK=True for param in kwargs: if kwargs[kwargs] is None: OK=False if not OK: #Print the list of Necessary params in kwargs return OK # the function can be called inside the classes as if check_params(Xs=Xs,Xt=Xt): do the stuff

What is nice is that we can give different keywords params depending on the class (semi supervised or else).

This function or at leats a clener version should be in utils.py

Sorry but this is not what you did in the last commit, you just added a condition, oi think the implementation above would be more future proof since it handles the warning/error message automatically, no need for else.

rflamary · 2017-08-28T07:51:15Z

ot/da.py

+            The transport source samples.
+        """
+
+        if np.array_equal(self.Xs, Xs):


self.Xs is Xs will test if the array is the same object and avoid the multiple tests and computation in np_array_equal.

…amples

rflamary

OK seems good to me.

Let's sleep on it and proceed with the merge tomorrow if everyone is OK with it.

Thank you to all of you and especially @Slasnista who did a wonderful job on these brand new classes.

After the merge we will wait for new features from @ncourty and release 0.4 on pipy.

agramfort · 2017-08-28T15:43:17Z

examples/da/plot_otda_color_images.py

@@ -61,62 +50,72 @@ def mat2im(X, shape):
 idx1 = np.random.randint(X1.shape[0], size=(nb,))
 idx2 = np.random.randint(X2.shape[0], size=(nb,))


use RandomState

You mean a seed ?

agramfort · 2017-08-28T15:44:35Z

examples/da/plot_otda_d2.py

+import ot
+
+# number of source and target points to generate
+ns = 150


n_samples_source = 150
n_samples_target = 150

it makes your comment above unnecessary and the code more readable

agramfort · 2017-08-28T15:45:24Z

examples/da/plot_otda_d2.py

+Xt, yt = ot.datasets.get_data_classif('3gauss2', nt)
+
+# Cost matrix
+M = ot.dist(Xs, Xt)


if it's doing euclidian distance than naming / API is not explicit enough IMO

dist is basically a wrapper around scipy cdist where you can choose the metric.

Also it shouldn't be computed here since it's not used anywhere it seems.

Actually I did compute it to plot it along with source and target points

agramfort · 2017-08-28T15:46:10Z

examples/da/plot_otda_mapping.py

+
+n = 100  # nb samples in source and target datasets
+theta = 2 * np.pi / 20
+nz = 0.1


nz ? could you use more explicit names?

agramfort · 2017-08-28T15:46:32Z

examples/da/plot_otda_mapping.py

+    max_iter=20, verbose=True)
+
+ot_mapping_linear.fit(
+    Xs=Xs, Xt=Xt)


it fits on one line

Slasnista · 2017-08-29T07:53:50Z

Final modifications according to @agramfort comments pushed

rflamary

Sorry last change and we can merge on my side I promise ;)

rflamary · 2017-08-29T11:18:46Z

examples/da/plot_otda_classes.py

 import ot

+np.random.seed(42)


Actually I think @agramfort prefer to use RandomSate as explained in
https://stackoverflow.com/questions/22994423/difference-between-np-random-seed-and-np-random-randomstate

This allow you to use controlable random generator without impacting the global one as is done in random.seed.

Note that we plan on doing proper dataset functions that accept a random state generator as parameter to allow full control (See #21 ).

no worries ;)

rflamary · 2017-08-29T12:13:22Z

OK it is merged, thank you to @ngayraud @Slasnista and @agramfort for the work.

Slasnista added 4 commits July 28, 2017 08:00

remove linewidth error message

553a456

first proposal for OT wrappers

ca9c9d6

small modifs according to NG proposals

84adadd

integrate AG comments

cd3397f

own BaseEstimator class written + rflamary comments addressed

bd7c7d2

rflamary reviewed Jul 31, 2017

View reviewed changes

ngayraud reviewed Jul 31, 2017

View reviewed changes

update SinkhornTransport class + added test for class

122b5bf

rflamary reviewed Aug 1, 2017

View reviewed changes

added EMDTransport Class from NG's code + added dedicated test

d9be6c2

Slasnista added 4 commits August 4, 2017 11:16

added test for fit_transform + correction of fit_transform bug (missi…

70be034

…ng return self)

added new class SinkhornLpl1Transport() + dedicated test

64880e7

added new class SinkhornL1l2Transport() + dedicated test

727077a

semi supervised mode supported

0b00590

added deprecation warning on old classes

0930223

solving log issues to avoid errors and adding further tests

2d4d0b4

Slasnista added 3 commits August 25, 2017 15:12

refactoring examples according to new DA classes

74ca2d7

small corrections for examples

f80693b

set properly path of data

892d7ce

rflamary reviewed Aug 28, 2017

View reviewed changes

Slasnista added 4 commits August 28, 2017 10:43

move no da objects into utils.py

55840f6

handling input arguments in fit, transform... methods + remove old ex…

a8fa91b

…amples

check input parameters with helper functions

c5d7c40

update readme

7d3fc95

rflamary approved these changes Aug 28, 2017

View reviewed changes

agramfort reviewed Aug 28, 2017

View reviewed changes

addressed AG comments + adding random seed

a29e22d

rflamary reviewed Aug 29, 2017

View reviewed changes

pass on examples | introduced RandomState

65de6fc

rflamary merged commit a2ec6e5 into PythonOT:master Aug 29, 2017

Slasnista mentioned this pull request Aug 30, 2017

Changes to LP solver: #25

Merged

		# ``if type(self).__module__.startswith('sklearn.')``.


		def distribution_estimation_uniform(X):

		@@ -706,12 +763,19 @@ def normalizeM(self, norm):
		self.M = np.log(1 + np.log(1 + self.M))


		@deprecated("The class OTDA_sinkhorn is deprecated in 0.3.1 and will be"


		class BaseTransport(BaseEstimator):

		def fit(self, Xs=None, ys=None, Xt=None, yt=None):


		return self

		def fit_transform(self, Xs=None, ys=None, Xt=None, yt=None):

		@@ -61,62 +50,72 @@ def mat2im(X, shape):
		idx1 = np.random.randint(X1.shape[0], size=(nb,))
		idx2 = np.random.randint(X2.shape[0], size=(nb,))

Domain adaptation Classes #22

Domain adaptation Classes #22

Uh oh!

Conversation

Slasnista commented Jul 28, 2017

Uh oh!

rflamary commented Jul 28, 2017

Uh oh!

rflamary commented Jul 31, 2017

Uh oh!

Slasnista commented Jul 31, 2017

Uh oh!

rflamary commented Jul 31, 2017

Uh oh!

Slasnista commented Jul 31, 2017

Uh oh!

rflamary left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Slasnista commented Aug 1, 2017

Uh oh!

rflamary left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ngayraud commented Aug 1, 2017

Uh oh!

Slasnista commented Aug 1, 2017

Uh oh!

rflamary commented Aug 1, 2017 via email

Uh oh!

Slasnista commented Aug 1, 2017

Uh oh!

ngayraud commented Aug 1, 2017

Uh oh!

Slasnista commented Aug 4, 2017

Uh oh!

rflamary commented Aug 4, 2017

Uh oh!

Slasnista commented Aug 4, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rflamary commented Aug 4, 2017

Uh oh!

Slasnista commented Aug 23, 2017

Uh oh!

rflamary commented Aug 25, 2017

Uh oh!

ncourty commented Aug 25, 2017

Uh oh!

agramfort commented Aug 25, 2017 via email

Uh oh!

Slasnista commented Aug 25, 2017

Uh oh!

rflamary commented Aug 25, 2017

Uh oh!

Slasnista commented Aug 25, 2017

Uh oh!

rflamary left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Slasnista commented Aug 4, 2017 •

edited

Loading

Slasnista Aug 28, 2017 •

edited

Loading