From 3b14418c77b73da5b4fb774a2f244ddcf59ee159 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" <gardener@tensorflow.org>
Date: Thu, 13 Oct 2016 12:04:59 -0800
Subject: [PATCH] Fix Beta, Wishart and Mixture sampling to use different seeds
per random draw. Change: 136075925
---
.../python/kernel_tests/beta_test.py | 17 ++++++++++++
.../python/kernel_tests/mixture_test.py | 26 +++++++++++++++++++
.../python/kernel_tests/student_t_test.py | 11 ++++----
.../python/kernel_tests/wishart_test.py | 24 +++++++++++++++++
.../contrib/distributions/python/ops/beta.py | 3 ++-
.../distributions/python/ops/mixture.py | 3 +++
.../distributions/python/ops/wishart.py | 4 ++-
7 files changed, 81 insertions(+), 7 deletions(-)
diff --git a/tensorflow/contrib/distributions/python/kernel_tests/beta_test.py b/tensorflow/contrib/distributions/python/kernel_tests/beta_test.py
index c4eea35dc12..e1a8a6d6025 100644
--- a/tensorflow/contrib/distributions/python/kernel_tests/beta_test.py
+++ b/tensorflow/contrib/distributions/python/kernel_tests/beta_test.py
@@ -245,6 +245,23 @@ class BetaTest(tf.test.TestCase):
stats.beta.var(a, b),
atol=1e-1)
+ # Test that sampling with the same seed twice gives the same results.
+ def testBetaSampleMultipleTimes(self):
+ with self.test_session():
+ a_val = 1.
+ b_val = 2.
+ n_val = 100
+
+ tf.set_random_seed(654321)
+ beta1 = tf.contrib.distributions.Beta(a=a_val, b=b_val, name="beta1")
+ samples1 = beta1.sample_n(n_val, seed=123456).eval()
+
+ tf.set_random_seed(654321)
+ beta2 = tf.contrib.distributions.Beta(a=a_val, b=b_val, name="beta2")
+ samples2 = beta2.sample_n(n_val, seed=123456).eval()
+
+ self.assertAllClose(samples1, samples2)
+
def testBetaSampleMultidimensional(self):
with self.test_session():
a = np.random.rand(3, 2, 2).astype(np.float32)
diff --git a/tensorflow/contrib/distributions/python/kernel_tests/mixture_test.py b/tensorflow/contrib/distributions/python/kernel_tests/mixture_test.py
index f871f1961c8..af15f36522a 100644
--- a/tensorflow/contrib/distributions/python/kernel_tests/mixture_test.py
+++ b/tensorflow/contrib/distributions/python/kernel_tests/mixture_test.py
@@ -334,6 +334,32 @@ class MixtureTest(tf.test.TestCase):
which_dist_samples = dist_sample_values[c][:size_c]
self.assertAllClose(which_dist_samples, sample_values[which_c])
+ # Test that sampling with the same seed twice gives the same results.
+ def testSampleMultipleTimes(self):
+ # 5 component mixture.
+ logits = [-10.0, -5.0, 0.0, 5.0, 10.0]
+ mus = [-5.0, 0.0, 5.0, 4.0, 20.0]
+ sigmas = [0.1, 5.0, 3.0, 0.2, 4.0]
+
+ with self.test_session():
+ n = 100
+
+ tf.set_random_seed(654321)
+ components = [distributions_py.Normal(
+ mu=mu, sigma=sigma) for mu, sigma in zip(mus, sigmas)]
+ cat = distributions_py.Categorical(logits, dtype=tf.int32, name="cat1")
+ dist1 = distributions_py.Mixture(cat, components, name="mixture1")
+ samples1 = dist1.sample_n(n, seed=123456).eval()
+
+ tf.set_random_seed(654321)
+ components2 = [distributions_py.Normal(
+ mu=mu, sigma=sigma) for mu, sigma in zip(mus, sigmas)]
+ cat2 = distributions_py.Categorical(logits, dtype=tf.int32, name="cat2")
+ dist2 = distributions_py.Mixture(cat2, components2, name="mixture2")
+ samples2 = dist2.sample_n(n, seed=123456).eval()
+
+ self.assertAllClose(samples1, samples2)
+
def testSampleScalarBatchMultivariate(self):
with self.test_session() as sess:
num_components = 3
diff --git a/tensorflow/contrib/distributions/python/kernel_tests/student_t_test.py b/tensorflow/contrib/distributions/python/kernel_tests/student_t_test.py
index 3ddd16d844a..c78ca2d6439 100644
--- a/tensorflow/contrib/distributions/python/kernel_tests/student_t_test.py
+++ b/tensorflow/contrib/distributions/python/kernel_tests/student_t_test.py
@@ -130,16 +130,17 @@ class StudentTTest(tf.test.TestCase):
mu_v = 3.0
sigma_v = np.sqrt(10.0)
n = tf.constant(100)
+
+ tf.set_random_seed(654321)
student = tf.contrib.distributions.StudentT(
- name="student_t1", df=df, mu=mu, sigma=sigma)
+ df=df, mu=mu, sigma=sigma, name="student_t1")
samples1 = student.sample_n(n, seed=123456).eval()
- # We need to start another test session, since this resets the internal
- # state.
- with self.test_session():
+ tf.set_random_seed(654321)
student2 = tf.contrib.distributions.StudentT(
- name="student_t2", df=df, mu=mu, sigma=sigma)
+ df=df, mu=mu, sigma=sigma, name="student_t2")
samples2 = student2.sample_n(n, seed=123456).eval()
+
self.assertAllClose(samples1, samples2)
def testStudentSampleSmallDfNoNan(self):
diff --git a/tensorflow/contrib/distributions/python/kernel_tests/wishart_test.py b/tensorflow/contrib/distributions/python/kernel_tests/wishart_test.py
index 521b0d4b2dd..8a0cf5ae1d9 100644
--- a/tensorflow/contrib/distributions/python/kernel_tests/wishart_test.py
+++ b/tensorflow/contrib/distributions/python/kernel_tests/wishart_test.py
@@ -149,6 +149,30 @@ class WishartCholeskyTest(tf.test.TestCase):
variance_estimate,
rtol=0.05)
+ # Test that sampling with the same seed twice gives the same results.
+ def testSampleMultipleTimes(self):
+ with self.test_session():
+ df = 4.
+ n_val = 100
+
+ tf.set_random_seed(654321)
+ chol_w1 = distributions.WishartCholesky(
+ df=df,
+ scale=chol(make_pd(1., 3)),
+ cholesky_input_output_matrices=False,
+ name="wishart1")
+ samples1 = chol_w1.sample_n(n_val, seed=123456).eval()
+
+ tf.set_random_seed(654321)
+ chol_w2 = distributions.WishartCholesky(
+ df=df,
+ scale=chol(make_pd(1., 3)),
+ cholesky_input_output_matrices=False,
+ name="wishart2")
+ samples2 = chol_w2.sample_n(n_val, seed=123456).eval()
+
+ self.assertAllClose(samples1, samples2)
+
def testProb(self):
with self.test_session():
# Generate some positive definite (pd) matrices and their Cholesky
diff --git a/tensorflow/contrib/distributions/python/ops/beta.py b/tensorflow/contrib/distributions/python/ops/beta.py
index 7f77254a644..684d6ec56b2 100644
--- a/tensorflow/contrib/distributions/python/ops/beta.py
+++ b/tensorflow/contrib/distributions/python/ops/beta.py
@@ -197,7 +197,8 @@ class Beta(distribution.Distribution):
gamma1_sample = random_ops.random_gamma(
[n,], a, dtype=self.dtype, seed=seed)
gamma2_sample = random_ops.random_gamma(
- [n,], b, dtype=self.dtype, seed=seed)
+ [n,], b, dtype=self.dtype,
+ seed=distribution_util.gen_new_seed(seed, "beta"))
beta_sample = gamma1_sample / (gamma1_sample + gamma2_sample)
return beta_sample
diff --git a/tensorflow/contrib/distributions/python/ops/mixture.py b/tensorflow/contrib/distributions/python/ops/mixture.py
index bd6f920c2cc..9827df6d10b 100644
--- a/tensorflow/contrib/distributions/python/ops/mixture.py
+++ b/tensorflow/contrib/distributions/python/ops/mixture.py
@@ -22,6 +22,7 @@ import numpy as np
from tensorflow.contrib.distributions.python.ops import categorical
from tensorflow.contrib.distributions.python.ops import distribution
+from tensorflow.contrib.distributions.python.ops import distribution_util
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_util
@@ -295,8 +296,10 @@ class Mixture(distribution.Distribution):
partitions=cat_samples,
num_partitions=self.num_components)
samples_class = [None for _ in range(self.num_components)]
+
for c in range(self.num_components):
n_class = array_ops.size(partitioned_samples_indices[c])
+ seed = distribution_util.gen_new_seed(seed, "mixture")
samples_class_c = self.components[c].sample_n(n_class, seed=seed)
# Pull out the correct batch entries from each index.
diff --git a/tensorflow/contrib/distributions/python/ops/wishart.py b/tensorflow/contrib/distributions/python/ops/wishart.py
index 92d952a1d24..29fa19d6fd1 100644
--- a/tensorflow/contrib/distributions/python/ops/wishart.py
+++ b/tensorflow/contrib/distributions/python/ops/wishart.py
@@ -22,6 +22,7 @@ import math
import numpy as np
from tensorflow.contrib.distributions.python.ops import distribution
+from tensorflow.contrib.distributions.python.ops import distribution_util
from tensorflow.contrib.distributions.python.ops import operator_pd_cholesky
from tensorflow.contrib.distributions.python.ops import operator_pd_full
from tensorflow.contrib.framework.python.framework import tensor_util as contrib_tensor_util
@@ -211,7 +212,8 @@ class _WishartOperatorPD(distribution.Distribution):
0.5 * self.df, self.dimension),
beta=0.5,
dtype=self.dtype,
- seed=seed)
+ seed=distribution_util.gen_new_seed(
+ seed, "wishart"))
# Complexity: O(nbk^2)
x = array_ops.matrix_band_part(x, -1, 0) # Tri-lower.