Merge remote-tracking branch 'upstream/master' into feature/typing-link

.gitignore

@@ -11,7 +11,6 @@ build
 .eggs/
 _readthedocs_build
 /TAGS
-/docs/source/reference/**/generated
 /tags
 chainer.egg-info/
 dist/
@@ -25,6 +24,4 @@ docs/my.state
 docs/my_mnist.model
 docs/mnist_result
 docs/*.png
-docs/source/reference/core
-docs/source/reference/generated
-docs/source/reference/util
+/docs/source/**/reference/**/generated

CODE_OF_CONDUCT.md

@@ -0,0 +1,7 @@
+# Chainer Code of Conduct
+
+Chainer follows the [NumFOCUS Code of Conduct][homepage] available at https://numfocus.org/code-of-conduct.
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be reported by contacting the project team at chainer@preferred.jp. 
+
+[homepage]: https://numfocus.org/

chainer/backends/cuda.py

@@ -312,7 +312,7 @@ def get_device(*args):
 
     .. note::
 
-        This API is deprecated. Please use
+        This API is deprecated since v3.0.0. Please use
         :func:`~chainer.backends.cuda.get_device_from_id`
         or :func:`~chainer.backends.cuda.get_device_from_array` instead.
 
@@ -716,6 +716,8 @@ def fuse(*args, **kwargs):
     """
     if available:
         return cupy.fuse(*args, **kwargs)
+    elif len(args) == 1 and len(kwargs) == 0 and callable(args[0]):
+        return args[0]
     else:
         return lambda f: f
 

chainer/dataset/convert.py

@@ -6,7 +6,6 @@
 import chainer
 from chainer import backend
 from chainer.backends import cuda
-from chainer import utils
 
 
 def to_device(device, x):
@@ -192,7 +191,7 @@ def _concat_arrays_with_padding(arrays, padding):
         for i in six.moves.range(len(arrays)):
             src = arrays[i]
             slices = tuple(slice(dim) for dim in src.shape)
-            utils._setitem(result, (i,) + slices, src)
+            result[(i,) + slices] = src
 
     return result
 

chainer/distributions/__init__.py

@@ -3,10 +3,12 @@
 from chainer.distributions.bernoulli import Bernoulli  # NOQA
 from chainer.distributions.beta import Beta  # NOQA
 from chainer.distributions.categorical import Categorical  # NOQA
+from chainer.distributions.cauchy import Cauchy  # 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
+from chainer.distributions.geometric import Geometric  # NOQA
 from chainer.distributions.gumbel import Gumbel  # NOQA
 from chainer.distributions.laplace import Laplace  # NOQA
 from chainer.distributions.log_normal import LogNormal  # NOQA

chainer/distributions/cauchy.py

@@ -0,0 +1,98 @@
+import warnings
+
+import numpy
+
+import chainer
+from chainer.backends import cuda
+from chainer import distribution
+from chainer.functions.math import exponential
+from chainer.functions.math import trigonometric
+
+
+def _cauchy_icdf(x):
+    x = chainer.as_variable(x)
+    h = (x - 0.5) * numpy.pi
+    y = chainer.functions.tan(h)
+    return y
+
+
+class Cauchy(distribution.Distribution):
+
+    """Cauchy Distribution.
+
+    The probability density function of the distribution is expressed as
+
+    .. math::
+        p(x;x_0,\\gamma) = \\frac{1}{\\pi}\\frac{\\gamma}{(x-x_0)^2+\\gamma^2}
+
+    Args:
+        loc(:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
+        :class:`cupy.ndarray`): Parameter of distribution representing the \
+        location :math:`\\x_0`.
+        scale(:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
+        :class:`cupy.ndarray`): Parameter of distribution representing the \
+        scale :math:`\\gamma`.
+    """
+
+    def __init__(self, loc, scale):
+        super(Cauchy, self).__init__()
+        self.loc = chainer.as_variable(loc)
+        self.scale = chainer.as_variable(scale)
+
+    @property
+    def batch_shape(self):
+        return self.loc.shape
+
+    def cdf(self, x):
+        return 1 / numpy.pi * trigonometric.arctan(
+            (x - self.loc) / self.scale) + 0.5
+
+    @property
+    def entropy(self):
+        return exponential.log(4 * numpy.pi * self.scale)
+
+    @property
+    def event_shape(self):
+        return ()
+
+    def icdf(self, x):
+        return self.loc + self.scale * _cauchy_icdf(x)
+
+    @property
+    def _is_gpu(self):
+        return isinstance(self.loc.data, cuda.ndarray)
+
+    def log_prob(self, x):
+        return - numpy.log(numpy.pi) + exponential.log(self.scale) \
+            - exponential.log((x - self.loc)**2 + self.scale**2)
+
+    @property
+    def mean(self):
+        warnings.warn("Mean of the cauchy distribution is undefined.",
+                      RuntimeWarning)
+        xp = cuda.get_array_module(self.loc)
+        return chainer.as_variable(xp.full_like(self.loc.data, xp.nan))
+
+    def sample_n(self, n):
+        xp = cuda.get_array_module(self.loc)
+        if xp is cuda.cupy:
+            eps = xp.random.standard_cauchy(
+                (n,)+self.loc.shape, dtype=self.loc.dtype)
+        else:
+            eps = xp.random.standard_cauchy(
+                (n,)+self.loc.shape).astype(self.loc.dtype)
+
+        noise = self.scale * eps + self.loc
+
+        return noise
+
+    @property
+    def support(self):
+        return 'real'
+
+    @property
+    def variance(self):
+        warnings.warn("Variance of the cauchy distribution is undefined.",
+                      RuntimeWarning)
+        xp = cuda.get_array_module(self.loc)
+        return chainer.as_variable(xp.full_like(self.loc.data, xp.nan))