# chainer/chainer

Merge 5b28e26 into c9fb8b8

ganow committed Apr 1, 2019
2 parents c9fb8b8 + 5b28e26 commit e47fbfdd2cb003535dc68af8d58d8d18f9ac762e
 @@ -10,6 +10,7 @@ from chainer.distributions.gamma import Gamma # NOQA from chainer.distributions.geometric import Geometric # NOQA from chainer.distributions.gumbel import Gumbel # NOQA from chainer.distributions.independent import Independent # NOQA from chainer.distributions.laplace import Laplace # NOQA from chainer.distributions.log_normal import LogNormal # NOQA from chainer.distributions.multivariate_normal import MultivariateNormal # NOQA
 @@ -131,6 +131,10 @@ def log_prob(self, x): def mean(self): return self.p @property def params(self): return {'logit': self.logit} def prob(self, x): x = chainer.as_variable(x) prob = x * self.p + (1 - x) * (1 - self.p)
 @@ -79,6 +79,10 @@ def log_prob(self, x): def mean(self): return self.a / self._a_plus_b @property def params(self): return {'a': self.a, 'b': self.b} def sample_n(self, n): xp = backend.get_array_module(self.a) eps = xp.random.beta(self.a.data, self.b.data, size=(n,)+self.a.shape)
 @@ -74,6 +74,10 @@ def log_prob(self, x): else: return self.log_p[mg + [x.astype(numpy.int32)]] @property def params(self): return {'p': self.p} def sample_n(self, n): xp = backend.get_array_module(self.p) onebyone_p = self.p.data.reshape(-1, self.p.shape[-1])
 @@ -80,6 +80,10 @@ def mean(self): xp = cuda.get_array_module(self.loc) return chainer.as_variable(xp.full_like(self.loc.data, xp.nan)) @property def params(self): return {'loc': self.loc, 'scale': self.scale} def sample_n(self, n): xp = cuda.get_array_module(self.loc) if xp is cuda.cupy:
 @@ -56,6 +56,10 @@ def log_prob(self, x): def mean(self): return self.k @property def params(self): return {'k': self.k} def sample_n(self, n): xp = cuda.get_array_module(self.k) if xp is cuda.cupy:
 @@ -68,6 +68,10 @@ def mean(self): alpha0 = expand_dims.expand_dims(self.alpha0, axis=-1) return self.alpha / alpha0 @property def params(self): return {'alpha': self.alpha} def sample_n(self, n): obo_alpha = self.alpha.data.reshape(-1, self.event_shape[0]) xp = cuda.get_array_module(self.alpha)
 @@ -69,6 +69,10 @@ def log_prob(self, x): def mean(self): return 1 / self.lam @property def params(self): return {'lam': self.lam} def sample_n(self, n): xp = cuda.get_array_module(self.lam) if xp is cuda.cupy:
 @@ -63,6 +63,10 @@ def log_prob(self, x): def mean(self): return self.k * self.theta @property def params(self): return {'k': self.k, 'theta': self.theta} def sample_n(self, n): xp = cuda.get_array_module(self.k) if xp is cuda.cupy:
 @@ -47,6 +47,10 @@ def log_prob(self, x): def mean(self): return 1 / self.p @property def params(self): return {'p': self.p} def sample_n(self, n): xp = cuda.get_array_module(self.p) if xp is cuda.cupy:
 @@ -71,6 +71,10 @@ def log_prob(self, x): def mean(self): return self.loc + EULER * self.scale @property def params(self): return {'loc': self.loc, 'scale': self.scale} def sample_n(self, n): xp = cuda.get_array_module(self.loc) if xp is cuda.cupy:
 @@ -0,0 +1,247 @@ import functools import operator import numpy from chainer.backend import cuda from chainer import distribution from chainer.functions.array import repeat from chainer.functions.array import reshape from chainer.functions.array import transpose from chainer.functions.math import sum as sum_mod from chainer.functions.math import prod from chainer.utils import cache class Independent(distribution.Distribution): """Independent distribution. Args: distribution (:class:`~chainer.Distribution`): The base distribution instance to transform. reinterpreted_batch_ndims (:class:`int`): Integer number of rightmost batch dims which will be regarded as event dims. When ``None`` all but the first batch axis (batch axis 0) will be transferred to event dimensions. """ def __init__(self, distribution, reinterpreted_batch_ndims=None): super(Independent, self).__init__() self.__distribution = distribution if reinterpreted_batch_ndims is None: reinterpreted_batch_ndims = \ self._get_default_reinterpreted_batch_ndims(distribution) elif reinterpreted_batch_ndims > len(distribution.batch_shape): raise ValueError( 'reinterpreted_batch_ndims must be less than or equal to the ' 'number of dimensions of `distribution.batch_shape`.') self.__reinterpreted_batch_ndims = reinterpreted_batch_ndims batch_ndim = \ len(self.distribution.batch_shape) - self.reinterpreted_batch_ndims self.__batch_shape = distribution.batch_shape[:batch_ndim] self.__event_shape = \ distribution.batch_shape[batch_ndim:] + distribution.event_shape @property def distribution(self): return self.__distribution @property def reinterpreted_batch_ndims(self): return self.__reinterpreted_batch_ndims @property def batch_shape(self): return self.__batch_shape @property def event_shape(self): return self.__event_shape @property def covariance(self): '''Returns the covariance of the distribution based on the original i.i.d. distribution. By definition, the covariance of the new distribution becomes block diagonal matrix. Let :math:`\\Sigma_{\\mathbf{x}}` be the covariance matrix of the original random variable :math:`\\mathbf{x} \\in \\mathbb{R}^d`, and :math:`\\mathbf{x}^{(1)}, \\mathbf{x}^{(2)}, \\cdots \\mathbf{x}^{(m)}` be the :math:`m` i.i.d. random variables, new covariance matrix :math:`\\Sigma_{\\mathbf{y}}` of :math:`\\mathbf{y} = [\\mathbf{x}^{(1)}, \\mathbf{x}^{(2)}, \\cdots, \\mathbf{x}^{(m)}] \\in \\mathbb{R}^{md}` can be written as .. math:: \\left[\\begin{array}{ccc} \\Sigma_{\\mathbf{x}^{1}} & & 0 \\\\ & \\ddots & \\\\ 0 & & \\Sigma_{\\mathbf{x}^{m}} \\end{array} \\right]. Note that this relationship holds only if the covariance matrix of the original distribution is given analytically. ''' num_repeat = functools.reduce( operator.mul, self.distribution.batch_shape[-self.reinterpreted_batch_ndims:], 1) dim = functools.reduce(operator.mul, self.distribution.event_shape, 1) cov = repeat.repeat( reshape.reshape( self.distribution.covariance, ((self.batch_shape) + (1, num_repeat, dim, dim))), num_repeat, axis=-4) cov = reshape.reshape( transpose.transpose( cov, axes=( tuple(range(len(self.batch_shape))) + (-4, -2, -3, -1))), self.batch_shape + (num_repeat * dim, num_repeat * dim)) block_indicator = self.xp.reshape( self._block_indicator, tuple([1] * len(self.batch_shape)) + self._block_indicator.shape) return cov * block_indicator @property def entropy(self): return self._reduce(sum_mod.sum, self.distribution.entropy) def cdf(self, x): return self._reduce(prod.prod, self.distribution.cdf(x)) def icdf(self, x): '''Cumulative distribution function for multivariate variable is not invertible. This function always raises :class:`RuntimeError`. Args: x (:class:`~chainer.Variable` or :ref:`ndarray`): Data points in the codomain of the distribution Raises: :class:`RuntimeError` ''' raise RuntimeError( 'Cumulative distribution function for multivariate variable ' 'is not invertible.') def log_cdf(self, x): return self._reduce(sum_mod.sum, self.distribution.log_cdf(x)) def log_prob(self, x): return self._reduce(sum_mod.sum, self.distribution.log_prob(x)) def log_survival_function(self, x): return self._reduce( sum_mod.sum, self.distribution.log_survival_function(x)) @property def mean(self): return self.distribution.mean @property def mode(self): return self.distribution.mode @property def params(self): return self.distribution.params def perplexity(self, x): return self._reduce(prod.prod, self.distribution.perplexity(x)) def prob(self, x): return self._reduce(prod.prod, self.distribution.prob(x)) def sample_n(self, n): return self.distribution.sample_n(n) @property def stddev(self): return self.distribution.stddev @property def support(self): return self.distribution.support def survival_function(self, x): return self._reduce(prod.prod, self.distribution.survival_function(x)) @property def variance(self): return self.distribution.variance @property def xp(self): return self.distribution.xp def _reduce(self, op, stat): range_ = tuple( (-1 - numpy.arange(self.reinterpreted_batch_ndims)).tolist()) return op(stat, axis=range_) def _get_default_reinterpreted_batch_ndims(self, distribution): ndims = len(distribution.batch_shape) return max(0, ndims - 1) @cache.cached_property def _block_indicator(self): num_repeat = functools.reduce( operator.mul, self.distribution.batch_shape[-self.reinterpreted_batch_ndims:], 1) dim = functools.reduce(operator.mul, self.distribution.event_shape, 1) block_indicator = numpy.fromfunction( lambda i, j: i // dim == j // dim, (num_repeat * dim, num_repeat * dim)).astype(int) if self.xp is cuda.cupy: block_indicator = cuda.to_gpu(block_indicator) return block_indicator @distribution.register_kl(Independent, Independent) def _kl_independent_independent(dist1, dist2): '''Batched KL divergence :math:`\\mathrm{KL}(\\mathrm{dist1} || \\mathrm{dist2})` for Independent distributions. We can leverage the fact that .. math:: \\mathrm{KL}( \\mathrm{Independent}(\\mathrm{dist1}) || \\mathrm{Independent}(\\mathrm{dist2})) = \\mathrm{sum}(\\mathrm{KL}(\\mathrm{dist1} || \\mathrm{dist2})) where the sum is over the ``reinterpreted_batch_ndims``. Args: dist1 (:class:`~chainer.distribution.Independent`): Instance of `Independent`. dist2 (:class:`~chainer.distribution.Independent`): Instance of `Independent`. Returns: Batchwise ``KL(dist1 || dist2)``. Raises: :class:`ValueError`: If the event space for ``dist1`` and ``dist2``, or their underlying distributions don't match. ''' p = dist1.distribution q = dist2.distribution # The KL between any two (non)-batched distributions is a scalar. # Given that the KL between two factored distributions is the sum, i.e. # KL(p1(x)p2(y) || q1(x)q2(y)) = KL(p1 || q1) + KL(q1 || q2), we compute # KL(p || q) and do a `reduce_sum` on the reinterpreted batch dimensions. if dist1.event_shape == dist2.event_shape: if p.event_shape == q.event_shape: num_reduce_dims = len(dist1.event_shape) - len(p.event_shape) reduce_dims = tuple([-i - 1 for i in range(0, num_reduce_dims)]) return sum_mod.sum( distribution.kl_divergence(p, q), axis=reduce_dims) else: raise NotImplementedError( 'KL between Independents with different ' 'event shapes not supported.') else: raise ValueError('Event shapes do not match.')
 @@ -115,6 +115,10 @@ def mean(self): def mode(self): return self.loc @property def params(self): return {'loc': self.loc, 'scale': self.scale} def prob(self, x): scale = self.scale return 0.5 / scale * exponential.exp(- abs(x - self.loc) / scale)
 @@ -65,6 +65,10 @@ def log_prob(self, x): def mean(self): return exponential.exp(self.mu + 0.5 * self.sigma ** 2) @property def params(self): return {'mu': self.mu, 'sigma': self.sigma} def sample_n(self, n): xp = backend.get_array_module(self.mu) if xp is cuda.cupy: