Merge pull request #5338 from YoshikawaMasashi/distributions/chisquare

add `D.Chisquare`

chainer/distributions/__init__.py

@@ -3,6 +3,7 @@
 from chainer.distributions.bernoulli import Bernoulli  # NOQA
 from chainer.distributions.beta import Beta  # NOQA
 from chainer.distributions.categorical import Categorical  # NOQA
+from chainer.distributions.chisquare import Chisquare  # NOQA
 from chainer.distributions.dirichlet import Dirichlet  # NOQA
 from chainer.distributions.exponential import Exponential  # NOQA
 from chainer.distributions.gamma import Gamma  # NOQA

chainer/distributions/chisquare.py

@@ -0,0 +1,71 @@
+import numpy
+
+import chainer
+from chainer.backends import cuda
+from chainer import distribution
+from chainer.functions.math import digamma
+from chainer.functions.math import exponential
+from chainer.functions.math import lgamma
+
+
+class Chisquare(distribution.Distribution):
+
+    """Chi-Square Distribution.
+
+    The probability density function of the distribution is expressed as
+
+    .. math::
+        p(x;k) = \\frac{1}{2^{k/2}\\Gamma(k/2)}x^{k/2-1}e^{-x/2}
+
+    Args:
+        k(:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
+        :class:`cupy.ndarray`): Parameter of distribution.
+    """
+
+    def __init__(self, k):
+        super(Chisquare, self).__init__()
+        self.__k = chainer.as_variable(k)
+
+    @property
+    def k(self):
+        return self.__k
+
+    @property
+    def batch_shape(self):
+        return self.k.shape
+
+    @property
+    def entropy(self):
+        return 0.5 * self.k + numpy.log(2.) + lgamma.lgamma(0.5 * self.k) \
+            + (1 - 0.5 * self.k) * digamma.digamma(0.5 * self.k)
+
+    @property
+    def event_shape(self):
+        return ()
+
+    def log_prob(self, x):
+        return - lgamma.lgamma(0.5 * self.k) - 0.5 * self.k * numpy.log(2.) \
+            + (0.5 * self.k - 1) * exponential.log(x) - 0.5 * x
+
+    @property
+    def mean(self):
+        return self.k
+
+    def sample_n(self, n):
+        xp = cuda.get_array_module(self.k)
+        if xp is cuda.cupy:
+            eps = xp.random.chisquare(
+                self.k.data, (n,)+self.k.shape, dtype=self.k.dtype)
+        else:
+            eps = xp.random.chisquare(
+                self.k.data, (n,)+self.k.shape).astype(self.k.dtype)
+        noise = chainer.Variable(eps)
+        return noise
+
+    @property
+    def support(self):
+        return 'positive'
+
+    @property
+    def variance(self):
+        return 2 * self.k

docs/source/reference/distributions.rst

@@ -17,6 +17,7 @@ Distributions
    chainer.distributions.Bernoulli
    chainer.distributions.Beta
    chainer.distributions.Categorical
+   chainer.distributions.Chisquare
    chainer.distributions.Dirichlet
    chainer.distributions.Exponential
    chainer.distributions.Gamma

tests/chainer_tests/distributions_tests/test_chisquare.py

@@ -0,0 +1,39 @@
+from chainer import distributions
+from chainer import testing
+import numpy
+
+
+@testing.parameterize(*testing.product({
+    'shape': [(2, 3), ()],
+    'is_variable': [True, False],
+    'sample_shape': [(3, 2), ()],
+}))
+@testing.fix_random()
+@testing.with_requires('scipy')
+class TestChisquare(testing.distribution_unittest):
+
+    scipy_onebyone = True
+
+    def setUp_configure(self):
+        from scipy import stats
+        self.dist = distributions.Chisquare
+        self.scipy_dist = stats.chi2
+
+        self.test_targets = set([
+            "batch_shape", "entropy", "event_shape", "log_prob", "mean",
+            "sample", "support", "variance"])
+
+        k = numpy.random.randint(1, 10, self.shape).astype(numpy.float32)
+        self.params = {"k": k}
+        self.scipy_params = {"df": k}
+
+        self.support = "positive"
+
+    def sample_for_test(self):
+        smp = numpy.random.chisquare(
+            df=1, size=self.sample_shape + self.shape
+        ).astype(numpy.float32)
+        return smp
+
+
+testing.run_module(__name__, __file__)