Remove legacy variational code, update to 0.3.1.

README.md

@@ -36,7 +36,7 @@ please clone the repository and run
 pip install .
 ```
 in the main directory. This will install ZhuSuan and its dependencies
-automatically. ZhuSuan also requires Tensorflow version 1.0 or later. Because
+automatically. ZhuSuan also requires **Tensorflow 1.4.0 or later**. Because
 users should choose whether to install the cpu or gpu version of Tensorflow,
 we do not include it in the dependencies. See
 [Installing Tensorflow](https://www.tensorflow.org/install/).
@@ -80,6 +80,8 @@ pip install ".[examples]"
   [HMC](examples/topic_models/lntm_mcem.py)
 * Probabilistic Matrix Factorization on MovieLens 1M:
   [HMC](examples/probabilistic_matrix_factorization/logistic_pmf_hmc.py)
+* Sparse Variational Gaussian Process:
+  [SGVB](examples/gaussian_process/svgp.py)
 
 ## Citing ZhuSuan
 

examples/gaussian_process/svgp.py

@@ -2,8 +2,7 @@
 # -*- coding: utf-8 -*-
 
 """
-Stochastic sparse variational Gaussian process (SVGP) with normal approximation
-for posterior.
+Stochastic variational inference for sparse Gaussian process (SVGP).
 
 For the formulation you can refer to, e.g., Section 2.1 of the following paper:
 

setup.py

@@ -6,23 +6,28 @@
 # To use a consistent encoding
 from codecs import open
 from os import path
+import re
 
 here = path.abspath(path.dirname(__file__))
 
+# Get the version string
+with open(path.join(here, 'zhusuan', '__init__.py'), encoding='utf-8') as f:
+    version = re.search(r'__version__ = \'(.*?)\'', f.read()).group(1)
+
 # Get the long description from the README file
 with open(path.join(here, 'README.md'), encoding='utf-8') as f:
-    long_description = f.read()
+    readme = f.read()
 
 setup(
     name='zhusuan',
 
     # Versions should comply with PEP440.  For a discussion on single-sourcing
     # the version across setup.py and the project code, see
     # https://packaging.python.org/en/latest/single_source_version.html
-    version='0.3.0',
+    version=version,
 
-    description='A Library for Generative Models',
-    long_description=long_description,
+    description='A Library for Bayesian Deep Learning',
+    long_description=readme,
 
     # The project's main homepage.
     url='https://github.com/thu-ml/zhusuan',

tests/distributions/test_multivariate.py

@@ -8,7 +8,8 @@
 from contextlib import contextmanager
 import tensorflow as tf
 import numpy as np
-from scipy import stats, misc, special
+from scipy import stats
+from scipy.special import logsumexp, factorial, gammaln
 
 from tests.distributions import utils
 from zhusuan.distributions.multivariate import *
@@ -230,11 +231,11 @@ def _test_value(logits, n_experiments, given, normalize_logits):
 
                 maybe_normalized_logits = logits
                 if normalize_logits:
-                    maybe_normalized_logits -= misc.logsumexp(
+                    maybe_normalized_logits -= logsumexp(
                         logits, axis=-1, keepdims=True)
                 n_experiments = np.sum(given, axis=-1)
-                target_log_p = np.log(misc.factorial(n_experiments)) - \
-                    np.sum(np.log(misc.factorial(given)), -1) + \
+                target_log_p = np.log(factorial(n_experiments)) - \
+                    np.sum(np.log(factorial(given)), -1) + \
                     np.sum(given * maybe_normalized_logits, -1)
                 self.assertAllClose(log_p.eval(), target_log_p)
                 p = dist.prob(given)
@@ -334,7 +335,7 @@ def _test_value(logits, given, normalize_logits):
 
                 maybe_normalized_logits = logits
                 if normalize_logits:
-                    maybe_normalized_logits -= misc.logsumexp(
+                    maybe_normalized_logits -= logsumexp(
                         logits, axis=-1, keepdims=True)
                 target_log_p = np.sum(given * maybe_normalized_logits, -1)
                 self.assertAllClose(log_p.eval(), target_log_p)
@@ -414,7 +415,7 @@ def test_value(self):
         with self.test_session(use_gpu=True):
             def _test_value(logits, given):
                 logits = np.array(logits, np.float32)
-                normalized_logits = logits - misc.logsumexp(
+                normalized_logits = logits - logsumexp(
                     logits, axis=-1, keepdims=True)
                 given = np.array(given, np.int32)
                 cat = OnehotCategorical(logits)
@@ -510,8 +511,7 @@ def dirichlet_logpdf(x, alpha):
             # scipy's implementation of dirichlet logpdf doesn't support
             # batch of x, we use this modified version.
             def _lnB(alpha):
-                return np.sum(special.gammaln(alpha)) - \
-                    special.gammaln(np.sum(alpha))
+                return np.sum(gammaln(alpha)) - gammaln(np.sum(alpha))
 
             lnB = _lnB(alpha)
             return - lnB + np.sum(np.log(x) * (alpha - 1), -1)
@@ -655,7 +655,7 @@ def _test_value(given, temperature, logits):
                 n = logits.shape[-1]
 
                 t = temperature
-                target_log_p = special.gammaln(n) + (n - 1) * np.log(t) + \
+                target_log_p = gammaln(n) + (n - 1) * np.log(t) + \
                     (logits - t * given).sum(axis=-1) - \
                     n * np.log(np.exp(logits - t * given).sum(axis=-1))
 
@@ -816,7 +816,7 @@ def _test_value(given, temperature, logits):
                 n = logits.shape[-1]
 
                 t = temperature
-                target_log_p = special.gammaln(n) + (n - 1) * np.log(t) + \
+                target_log_p = gammaln(n) + (n - 1) * np.log(t) + \
                     (logits - (t + 1) * np.log(given)).sum(axis=-1) - \
                     n * np.log(np.exp(logits - t * np.log(given)).sum(axis=-1))
 

tests/distributions/test_univariate.py

@@ -7,7 +7,8 @@
 
 import tensorflow as tf
 import numpy as np
-from scipy import stats, misc
+from scipy import stats
+from scipy.special import logsumexp
 
 from tests.distributions import utils
 from zhusuan.distributions.univariate import *
@@ -537,7 +538,7 @@ def test_value(self):
         with self.test_session(use_gpu=True):
             def _test_value(logits, given):
                 logits = np.array(logits, np.float32)
-                normalized_logits = logits - misc.logsumexp(
+                normalized_logits = logits - logsumexp(
                     logits, axis=-1, keepdims=True)
                 given = np.array(given, np.int32)
                 cat = Categorical(logits)

tests/distributions/test_utils.py

@@ -7,7 +7,7 @@
 
 import tensorflow as tf
 import numpy as np
-from scipy import misc
+from scipy.special import factorial
 
 from zhusuan.distributions.utils import *
 from zhusuan.distributions.utils import get_shape_list, get_shape_at, \
@@ -20,8 +20,8 @@ def test_log_combination(self):
             def _test_func(n, ks):
                 tf_n = tf.convert_to_tensor(n, tf.float32)
                 tf_ks = tf.convert_to_tensor(ks, tf.float32)
-                true_value = np.log(misc.factorial(n)) - \
-                    np.sum(np.log(misc.factorial(ks)), axis=-1)
+                true_value = np.log(factorial(n)) - \
+                    np.sum(np.log(factorial(ks)), axis=-1)
                 test_value = log_combination(tf_n, tf_ks).eval()
                 self.assertAllClose(true_value, test_value)
 

tests/test_utils.py

@@ -5,7 +5,7 @@
 from __future__ import division
 
 import tensorflow as tf
-from scipy import misc
+from scipy.special import logsumexp
 import numpy as np
 import six
 
@@ -254,7 +254,7 @@ def test_log_sum_exp(self):
             a = np.array([[[1., 3., 0.2], [0.7, 2., 1e-6]],
                           [[0., 1e6, 1.], [1., 1., 1.]]])
             for keepdims in [True, False]:
-                true_values = misc.logsumexp(a, (0, 2), keepdims=keepdims)
+                true_values = logsumexp(a, (0, 2), keepdims=keepdims)
                 test_values = sess.run(log_sum_exp(
                     tf.constant(a), (0, 2), keepdims))
                 self.assertAllClose(test_values, true_values)
@@ -267,7 +267,7 @@ def test_log_mean_exp(self):
             a = np.array([[[1., 3., 0.2], [0.7, 2., 1e-6]],
                           [[0., 1e6, 1.], [1., 1., 1.]]])
             for keepdims in [True, False]:
-                true_values = misc.logsumexp(a, (0, 2), keepdims=keepdims) - \
+                true_values = logsumexp(a, (0, 2), keepdims=keepdims) - \
                               np.log(a.shape[0] * a.shape[2])
                 test_values = sess.run(log_mean_exp(
                     tf.constant(a), (0, 2), keepdims))

zhusuan/__init__.py

@@ -10,4 +10,5 @@
 from .utils import *
 from .legacy import *
 
-__version__ = '0.3.0'
+
+__version__ = '0.3.1'

zhusuan/legacy/__init__.py

@@ -1,4 +1,2 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
-
-from .variational_legacy import *