Skip to content

Commit 2480999

Browse files
antoinedemathelinantoinedemathelin
authored
Merge pull request #82 from antoinedemathelin/master
Add IWC
2 parents 42cbd04 + af60eb2 commit 2480999

File tree

3 files changed

+249
-1
lines changed

3 files changed

+249
-1
lines changed

‎adapt/instance_based/__init__.py‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,8 @@
1212
from ._iwn import IWN
1313
from ._ulsif import ULSIF
1414
from ._rulsif import RULSIF
15+
from ._iwc import IWC
1516

1617
__all__ = ["LDM", "KLIEP", "KMM", "TrAdaBoost", "TrAdaBoostR2",
1718
"TwoStageTrAdaBoostR2", "WANN", "NearestNeighborsWeighting",
18-
"BalancedWeighting", "IWN", "ULSIF", "RULSIF"]
19+
"BalancedWeighting", "IWN", "ULSIF", "RULSIF", "IWC"]

‎adapt/instance_based/_iwc.py‎

Lines changed: 213 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,213 @@
1+
"""
2+
IWC
3+
"""
4+
5+
import inspect
6+
7+
import numpy as np
8+
from sklearn.utils import check_array
9+
from sklearn.linear_model import LogisticRegression
10+
from sklearn.base import BaseEstimator
11+
from sklearn.exceptions import NotFittedError
12+
13+
from adapt.base import BaseAdaptEstimator, make_insert_doc
14+
from adapt.utils import check_arrays, set_random_seed, check_estimator
15+
16+
EPS = np.finfo(float).eps
17+
18+
19+
@make_insert_doc()
20+
class IWC(BaseAdaptEstimator):
21+
"""
22+
IWC: Importance Weighting Classifier
23+
24+
Importance weighting based on the output of a domain classifier
25+
which discriminates between source and target data.
26+
27+
The source importance weighting are given with the following formula:
28+
29+
.. math::
30+
31+
w(x) = \frac{1}{P(x \in Source)} - 1
32+
33+
Parameters
34+
----------
35+
classifier : object (default=None)
36+
Binary classifier trained to discriminate
37+
between source and target data.
38+
39+
cl_params : dict (default=None)
40+
Dictionnary of parameters that will
41+
be given in the `fit` and/or `compile` methods
42+
of the classifier.
43+
44+
Attributes
45+
----------
46+
classifier_ : object
47+
Fitted classifier.
48+
49+
estimator_ : object
50+
Fitted estimator.
51+
52+
See also
53+
--------
54+
NearestNeighborsWeighting
55+
IWN
56+
57+
Examples
58+
--------
59+
>>> from sklearn.linear_model import RidgeClassifier
60+
>>> from adapt.utils import make_classification_da
61+
>>> from adapt.instance_based import IWC
62+
>>> Xs, ys, Xt, yt = make_classification_da()
63+
>>> model = IWC(RidgeClassifier(0.), classifier=RidgeClassifier(0.),
64+
... Xt=Xt, random_state=0)
65+
>>> model.fit(Xs, ys);
66+
>>> model.score(Xt, yt)
67+
0.74
68+
69+
References
70+
----------
71+
.. [1] `[1] <https://icml.cc/imls/conferences/2007/proceedings/papers/303.pdf>`_ \
72+
Steffen Bickel, Michael Bruckner, Tobias Scheffer. "Discriminative Learning for Differing \
73+
Training and Test Distributions". In ICML 2007
74+
"""
75+
def __init__(self,
76+
estimator=None,
77+
Xt=None,
78+
yt=None,
79+
classifier=None,
80+
cl_params=None,
81+
copy=True,
82+
verbose=1,
83+
random_state=None,
84+
**params):
85+
86+
names = self._get_param_names()
87+
kwargs = {k: v for k, v in locals().items() if k in names}
88+
kwargs.update(params)
89+
super().__init__(**kwargs)
90+
91+
92+
def fit_weights(self, Xs, Xt, warm_start=False, **kwargs):
93+
"""
94+
Fit importance weighting.
95+
96+
Parameters
97+
----------
98+
Xs : array
99+
Input source data.
100+
101+
Xt : array
102+
Input target data.
103+
104+
warm_start : bool (default=False)
105+
Weither to train the domain classifier
106+
from scratch or not.
107+
If False, the classifier is trained from scratch.
108+
109+
kwargs : key, value argument
110+
Not used, present here for adapt consistency.
111+
112+
Returns
113+
-------
114+
weights_ : sample weights
115+
"""
116+
Xs = check_array(Xs)
117+
Xt = check_array(Xt)
118+
set_random_seed(self.random_state)
119+
120+
if self.cl_params is None:
121+
self.cl_params_ = {}
122+
else:
123+
self.cl_params_ = self.cl_params
124+
125+
if (not warm_start) or (not hasattr(self, "classifier_")):
126+
if self.classifier is None:
127+
self.classifier_ = LogisticRegression(penalty="none")
128+
else:
129+
self.classifier_ = check_estimator(self.classifier,
130+
copy=True,
131+
force_copy=True)
132+
133+
if hasattr(self.classifier_, "compile"):
134+
args = [
135+
p.name
136+
for p in inspect.signature(self.classifier_.compile).parameters.values()
137+
if p.name != "self" and p.kind != p.VAR_KEYWORD
138+
]
139+
compile_params = {}
140+
for key, value in self.cl_params_.items():
141+
if key in args:
142+
compile_params[key] = value
143+
self.classifier_.compile(**compile_params)
144+
145+
args = [
146+
p.name
147+
for p in inspect.signature(self.classifier_.fit).parameters.values()
148+
if p.name != "self" and p.kind != p.VAR_KEYWORD
149+
]
150+
fit_params = {}
151+
for key, value in self.cl_params_.items():
152+
if key in args:
153+
fit_params[key] = value
154+
155+
X = np.concatenate((Xs, Xt))
156+
y = np.concatenate((np.ones(Xs.shape[0]), np.zeros(Xt.shape[0])))
157+
shuffle_index = np.random.choice(len(X), len(X), replace=False)
158+
X = X[shuffle_index]
159+
y = y[shuffle_index]
160+
161+
self.classifier_.fit(X, y, **fit_params)
162+
163+
if isinstance(self.classifier_, BaseEstimator):
164+
if hasattr(self.classifier_, "predict_proba"):
165+
y_pred = self.classifier_.predict_proba(Xs)[:, 1]
166+
elif hasattr(self.classifier_, "_predict_proba_lr"):
167+
y_pred = self.classifier_._predict_proba_lr(Xs)[:, 1]
168+
else:
169+
y_pred = self.classifier_.predict(Xs).ravel()
170+
else:
171+
y_pred = self.classifier_.predict(Xs).ravel()
172+
173+
self.weights_ = 1. / (y_pred + EPS) - 1.
174+
175+
return self.weights_
176+
177+
178+
def predict_weights(self, X=None):
179+
"""
180+
Return fitted source weights
181+
182+
If ``None``, the fitted source weights are returned.
183+
Else, sample weights are computing using the fitted
184+
``classifier_``.
185+
186+
Parameters
187+
----------
188+
X : array (default=None)
189+
Input data.
190+
191+
Returns
192+
-------
193+
weights_ : sample weights
194+
"""
195+
if hasattr(self, "weights_"):
196+
if X is None:
197+
return self.weights_
198+
else:
199+
X = check_array(X)
200+
if isinstance(self.classifier_, BaseEstimator):
201+
if hasattr(self.classifier_, "predict_proba"):
202+
y_pred = self.classifier_.predict_proba(X)[:, 1]
203+
elif hasattr(self.classifier_, "_predict_proba_lr"):
204+
y_pred = self.classifier_._predict_proba_lr(X)[:, 1]
205+
else:
206+
y_pred = self.classifier_.predict(X).ravel()
207+
else:
208+
y_pred = self.classifier_.predict(X).ravel()
209+
weights = 1. / (y_pred + EPS) - 1.
210+
return weights
211+
else:
212+
raise NotFittedError("Weights are not fitted yet, please "
213+
"call 'fit_weights' or 'fit' first.")

‎tests/test_iwc.py‎

Lines changed: 34 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,34 @@
1+
"""
2+
Test functions for iwn module.
3+
"""
4+
5+
from sklearn.linear_model import RidgeClassifier
6+
from adapt.utils import make_classification_da
7+
from adapt.instance_based import IWC
8+
from adapt.utils import get_default_discriminator
9+
10+
Xs, ys, Xt, yt = make_classification_da()
11+
12+
def test_iwn():
13+
model = IWC(RidgeClassifier(0.), classifier=RidgeClassifier(0.),
14+
Xt=Xt, random_state=0)
15+
model.fit(Xs, ys);
16+
model.predict(Xt)
17+
model.score(Xt, yt)
18+
19+
20+
def test_default_classif():
21+
model = IWC(RidgeClassifier(0.), classifier=None,
22+
Xt=Xt, random_state=0)
23+
model.fit(Xs, ys);
24+
model.predict(Xt)
25+
model.score(Xt, yt)
26+
27+
28+
def test_nn_classif():
29+
model = IWC(RidgeClassifier(0.), classifier=get_default_discriminator(),
30+
cl_params=dict(epochs=10, optimizer="adam", loss="bce", verbose=0),
31+
Xt=Xt, random_state=0)
32+
model.fit(Xs, ys);
33+
model.predict(Xt)
34+
model.score(Xt, yt)

0 commit comments

Comments
 (0)