Change tfp/experimental to use internal imports to speed up tests.

  - Updates a few tests to avoid importing `distributions` or `bijectors` wholesale.

PiperOrigin-RevId: 472567989

tensorflow_probability/python/distributions/BUILD

@@ -766,7 +766,7 @@ multi_substrate_py_library(
         "//tensorflow_probability/python/internal:distribution_util",
         "//tensorflow_probability/python/internal:dtype_util",
         "//tensorflow_probability/python/internal:tensorshape_util",
-        "//tensorflow_probability/python/math/psd_kernels",
+        "//tensorflow_probability/python/math/psd_kernels:schur_complement",
         "//tensorflow_probability/python/util",
     ],
 )
@@ -1025,6 +1025,8 @@ multi_substrate_py_library(
     srcs = ["inverse_gaussian.py"],
     deps = [
         ":distribution",
+        ":inflated",
+        ":negative_binomial",
         ":normal",
         # numpy dep,
         # tensorflow dep,
@@ -1145,6 +1147,8 @@ multi_substrate_py_library(
     name = "joint_distribution_util",
     srcs = ["joint_distribution_util.py"],
     deps = [
+        ":independent",
+        ":joint_distribution_auto_batched",
         ":joint_distribution_named",
         ":joint_distribution_sequential",
         # tensorflow dep,
@@ -1420,7 +1424,6 @@ multi_substrate_py_library(
         "//tensorflow_probability/python/internal:parameter_properties",
         "//tensorflow_probability/python/internal:reparameterization",
         "//tensorflow_probability/python/internal:tensor_util",
-        "//tensorflow_probability/python/math",
         "//tensorflow_probability/python/util:seed_stream",
     ],
 )
@@ -2156,6 +2159,8 @@ multi_substrate_py_library(
         "//tensorflow_probability/python/internal:reparameterization",
         "//tensorflow_probability/python/internal:tensor_util",
         "//tensorflow_probability/python/internal:tensorshape_util",
+        "//tensorflow_probability/python/math/psd_kernels:positive_semidefinite_kernel",
+        "//tensorflow_probability/python/math/psd_kernels:schur_complement",
     ],
 )
 
@@ -2470,7 +2475,7 @@ multi_substrate_py_test(
         "//tensorflow_probability/python/bijectors:scale_matvec_tril",
         "//tensorflow_probability/python/internal:reparameterization",
         "//tensorflow_probability/python/internal:test_util",
-        "//tensorflow_probability/python/math",
+        "//tensorflow_probability/python/math:linalg",
     ],
 )
 
@@ -3198,8 +3203,12 @@ multi_substrate_py_test(
     name = "inflated_test",
     srcs = ["inflated_test.py"],
     deps = [
+        ":inflated",
+        ":negative_binomial",
+        ":normal",
         # numpy dep,
-        "//tensorflow_probability",
+        # tensorflow dep,
+        "//tensorflow_probability/python/experimental/util",
         "//tensorflow_probability/python/internal:test_util",
     ],
 )
@@ -3230,7 +3239,7 @@ multi_substrate_py_test(
         # scipy dep,
         # tensorflow dep,
         "//tensorflow_probability/python/internal:test_util",
-        "//tensorflow_probability/python/math",
+        "//tensorflow_probability/python/math:gradient",
     ],
 )
 

tensorflow_probability/python/distributions/autoregressive_test.py

@@ -20,7 +20,6 @@
 
 import numpy as np
 import tensorflow.compat.v2 as tf
-from tensorflow_probability.python import math
 from tensorflow_probability.python.bijectors import masked_autoregressive
 from tensorflow_probability.python.bijectors import scale_matvec_tril
 from tensorflow_probability.python.distributions import autoregressive
@@ -32,6 +31,7 @@
 from tensorflow_probability.python.distributions import transformed_distribution
 from tensorflow_probability.python.internal import reparameterization
 from tensorflow_probability.python.internal import test_util
+from tensorflow_probability.python.math import linalg
 
 
 @test_util.test_all_tf_execution_regimes
@@ -46,7 +46,7 @@ def setUp(self):
   def _random_scale_tril(self, event_size):
     n = np.int32(event_size * (event_size + 1) // 2)
     p = 2. * self._rng.random_sample(n).astype(np.float32) - 1.
-    return math.fill_triangular(0.25 * p)
+    return linalg.fill_triangular(0.25 * p)
 
   def _normal_fn(self, affine_bijector):
     def _fn(samples):

tensorflow_probability/python/distributions/gaussian_process_regression_model.py

@@ -25,7 +25,7 @@
 from tensorflow_probability.python.internal import parameter_properties
 from tensorflow_probability.python.internal import tensor_util
 from tensorflow_probability.python.internal import tensorshape_util
-from tensorflow_probability.python.math import psd_kernels as tfpk
+from tensorflow_probability.python.math.psd_kernels import schur_complement
 from tensorflow.python.util import deprecation  # pylint: disable=g-direct-tensorflow-import
 
 
@@ -534,7 +534,7 @@ def __init__(self,
 
       with tf.name_scope('init'):
         if _conditional_kernel is None:
-          _conditional_kernel = tfpk.SchurComplement(
+          _conditional_kernel = schur_complement.SchurComplement(
               base_kernel=kernel,
               fixed_inputs=observation_index_points,
               cholesky_fn=cholesky_fn,
@@ -749,7 +749,7 @@ def precompute_regression_model(
       if cholesky_fn is None:
         cholesky_fn = cholesky_util.make_cholesky_with_jitter_fn(jitter)
 
-      conditional_kernel = tfpk.SchurComplement.with_precomputed_divisor(
+      conditional_kernel = schur_complement.SchurComplement.with_precomputed_divisor(
           base_kernel=kernel,
           fixed_inputs=observation_index_points,
           fixed_inputs_is_missing=observations_is_missing,

tensorflow_probability/python/distributions/inflated_test.py

@@ -16,41 +16,42 @@
 import numpy as np
 import tensorflow.compat.v2 as tf
 
-import tensorflow_probability as tfp
-
-from tensorflow_probability.python import distributions as tfd
-from tensorflow_probability.python import experimental as tfe
+from tensorflow_probability.python.distributions import inflated
+from tensorflow_probability.python.distributions import negative_binomial
+from tensorflow_probability.python.distributions import normal
+from tensorflow_probability.python.experimental import util
+from tensorflow_probability.python.experimental.util import trainable
 from tensorflow_probability.python.internal import test_util
 
-tfe_util = tfp.experimental.util
-
 
 class DistributionsTest(test_util.TestCase):
 
   def test_inflated(self):
-    zinb = tfd.Inflated(
-        tfd.NegativeBinomial(5.0, probs=0.1), inflated_loc_probs=0.2)
+    zinb = inflated.Inflated(
+        negative_binomial.NegativeBinomial(5.0, probs=0.1),
+        inflated_loc_probs=0.2)
     samples = zinb.sample(sample_shape=10, seed=test_util.test_seed())
     self.assertEqual((10,), samples.shape)
 
-    spike_and_slab = tfd.Inflated(
-        tfd.Normal(loc=1.0, scale=2.0), inflated_loc_probs=0.5)
+    spike_and_slab = inflated.Inflated(
+        normal.Normal(loc=1.0, scale=2.0), inflated_loc_probs=0.5)
     lprob = self.evaluate(spike_and_slab.log_prob(99.0))
     self.assertLess(lprob, 0.0)
 
-    param_props = tfd.Inflated.parameter_properties(dtype=tf.float32)
+    param_props = inflated.Inflated.parameter_properties(dtype=tf.float32)
     self.assertFalse(param_props['distribution'].is_tensor)
     self.assertTrue(param_props['inflated_loc_logits'].is_preferred)
     self.assertFalse(param_props['inflated_loc_probs'].is_preferred)
     self.assertTrue(param_props['inflated_loc'].is_tensor)
 
   def test_inflated_batched(self):
-    nb = tfd.NegativeBinomial(
+    nb = negative_binomial.NegativeBinomial(
         total_count=np.array([1.0, 2.0, 3.0, 4.0, 5.0], dtype=np.float32),
         logits=np.array([-2.0, -1.0, 0.0, 1.0, 2.0], dtype=np.float32))
-    zinb = tfd.Inflated(
-        nb, inflated_loc_probs=np.array(
-            [0.2, 0.4, 0.6, 0.8, 1.0], dtype=np.float32))
+    zinb = inflated.Inflated(
+        nb,
+        inflated_loc_probs=np.array([0.2, 0.4, 0.6, 0.8, 1.0],
+                                    dtype=np.float32))
 
     lprob = zinb.log_prob([0, 1, 2, 3, 4])
     self.assertEqual((5,), lprob.shape)
@@ -59,24 +60,24 @@ def test_inflated_batched(self):
     self.assertEqual((5,), samples.shape)
 
   def test_inflated_factory(self):
-    spike_and_slab_class = tfe.distributions.inflated_factory(
-        'SpikeAndSlab', tfd.Normal, 0.0)
+    spike_and_slab_class = inflated.inflated_factory('SpikeAndSlab',
+                                                     normal.Normal, 0.0)
     spike_and_slab = spike_and_slab_class(
         inflated_loc_probs=0.3, loc=5.0, scale=2.0)
-    spike_and_slab2 = tfd.Inflated(
-        tfd.Normal(loc=5.0, scale=2.0), inflated_loc_probs=0.3)
+    spike_and_slab2 = inflated.Inflated(
+        normal.Normal(loc=5.0, scale=2.0), inflated_loc_probs=0.3)
     self.assertEqual(
         self.evaluate(spike_and_slab.log_prob(7.0)),
         self.evaluate(spike_and_slab2.log_prob(7.0)))
 
   def test_zero_inflated_negative_binomial(self):
-    zinb = tfd.ZeroInflatedNegativeBinomial(
+    zinb = inflated.ZeroInflatedNegativeBinomial(
         inflated_loc_probs=0.2, probs=0.5, total_count=10.0)
     self.assertEqual('ZeroInflatedNegativeBinomial', zinb.name)
 
   def test_zinb_is_trainable(self):
-    init_fn, apply_fn = tfe_util.make_trainable_stateless(
-        tfd.ZeroInflatedNegativeBinomial,
+    init_fn, apply_fn = trainable.make_trainable_stateless(
+        inflated.ZeroInflatedNegativeBinomial,
         batch_and_event_shape=[5],
         parameter_dtype=tf.float32)
     init_obj = init_fn(seed=test_util.test_seed())
@@ -96,8 +97,9 @@ def test_zinb_is_trainable(self):
       disable_numpy=True,
       reason='Only TF has composite tensors')
   def test_zinb_as_composite_tensor(self):
-    zinb = tfd.ZeroInflatedNegativeBinomial(0.1, total_count=10.0, probs=0.4)
-    comp_zinb = tfe.as_composite(zinb)
+    zinb = inflated.ZeroInflatedNegativeBinomial(
+        0.1, total_count=10.0, probs=0.4)
+    comp_zinb = util.as_composite(zinb)
     unused_as_tensors = tf.nest.flatten(comp_zinb)
 
 

tensorflow_probability/python/distributions/inverse_gaussian_test.py

@@ -17,9 +17,9 @@
 from scipy import stats
 import tensorflow.compat.v1 as tf1
 import tensorflow.compat.v2 as tf
-from tensorflow_probability.python import math as tfm
 from tensorflow_probability.python.distributions import inverse_gaussian
 from tensorflow_probability.python.internal import test_util
+from tensorflow_probability.python.math import gradient
 
 
 def _scipy_invgauss(loc, concentration):
@@ -373,7 +373,7 @@ def testInverseGaussianSampleMultidimensionalVariance(self):
   def testInverseGaussianFullyReparameterized(self):
     concentration = tf.constant(4.0)
     loc = tf.constant(3.0)
-    _, [grad_concentration, grad_loc] = tfm.value_and_gradient(
+    _, [grad_concentration, grad_loc] = gradient.value_and_gradient(
         lambda a, b: inverse_gaussian.InverseGaussian(a, b, validate_args=True).  # pylint: disable=g-long-lambda
         sample(100, seed=test_util.test_seed()),
         [concentration, loc])
@@ -393,7 +393,7 @@ def gen_samples(l, c):
           2, seed=test_util.test_seed())
 
     samples, [loc_grad, concentration_grad] = self.evaluate(
-        tfm.value_and_gradient(gen_samples, [loc, concentration]))
+        gradient.value_and_gradient(gen_samples, [loc, concentration]))
     self.assertEqual(samples.shape, (2, 4, 3))
     self.assertEqual(concentration_grad.shape, concentration.shape)
     self.assertEqual(loc_grad.shape, loc.shape)

tensorflow_probability/python/distributions/student_t_process_regression_model.py

@@ -28,7 +28,8 @@
 from tensorflow_probability.python.internal import prefer_static as ps
 from tensorflow_probability.python.internal import tensor_util
 from tensorflow_probability.python.internal import tensorshape_util
-from tensorflow_probability.python.math import psd_kernels as tfpk
+from tensorflow_probability.python.math.psd_kernels import positive_semidefinite_kernel as psd_kernel
+from tensorflow_probability.python.math.psd_kernels import schur_complement as schur_complement_lib
 
 
 __all__ = [
@@ -102,7 +103,7 @@ def _validate_observation_data(
               index_point_count, observation_count))
 
 
-class DampedSchurComplement(tfpk.AutoCompositeTensorPsdKernel):
+class DampedSchurComplement(psd_kernel.AutoCompositeTensorPsdKernel):
   """Schur complement kernel, damped by scalar factors.
 
   This kernel is the same as the SchurComplement kernel, except we multiply by
@@ -398,7 +399,7 @@ def __init__(
         if _conditional_kernel is None:
           _conditional_kernel = DampedSchurComplement(
               df=df,
-              schur_complement=tfpk.SchurComplement(
+              schur_complement=schur_complement_lib.SchurComplement(
                   base_kernel=kernel,
                   fixed_inputs=self._observation_index_points,
                   diag_shift=observation_noise_variance),
@@ -606,7 +607,7 @@ def precompute_regression_model(
 
       conditional_kernel = DampedSchurComplement(
           df=df,
-          schur_complement=tfpk.SchurComplement(
+          schur_complement=schur_complement_lib.SchurComplement(
               base_kernel=kernel,
               fixed_inputs=observation_index_points,
               diag_shift=observation_noise_variance),