def disable_v2_behavior():
"""Disables TensorFlow 2.x behaviors.
This function can be called at the beginning of the program (before `Tensors`,
`Graphs` or other structures have been created, and before devices have been
initialized. It switches all global behaviors that are different between
TensorFlow 1.x and 2.x to behave as intended for 1.x.
User can call this function to disable 2.x behavior during complex migrations.
"""
tf2.disable()
ops.disable_eager_execution()
tensor_shape.disable_v2_tensorshape() # Also switched by tf2
variable_scope.disable_resource_variables()
ops.disable_tensor_equality()
# Disables TensorArrayV2 and control flow V2.
control_flow_v2_toggles.disable_control_flow_v2()
# Make sure internal uses of tf.data symbols map to V1 versions.
dataset_ops.Dataset = dataset_ops.DatasetV1
readers.FixedLengthRecordDataset = readers.FixedLengthRecordDatasetV1
readers.TFRecordDataset = readers.TFRecordDatasetV1
readers.TextLineDataset = readers.TextLineDatasetV1
counter.Counter = counter.CounterV1
interleave_ops.choose_from_datasets = interleave_ops.choose_from_datasets_v1
interleave_ops.sample_from_datasets = interleave_ops.sample_from_datasets_v1
random_ops.RandomDataset = random_ops.RandomDatasetV1
exp_readers.CsvDataset = exp_readers.CsvDatasetV1
exp_readers.SqlDataset = exp_readers.SqlDatasetV1
exp_readers.make_batched_features_dataset = (
exp_readers.make_batched_features_dataset_v1)
exp_readers.make_csv_dataset = exp_readers.make_csv_dataset_v1
def testEqualityBroadcast(self):
default = ops.Tensor._USE_EQUALITY
try:
tf_a = constant_op.constant([1, 1])
tf_b = constant_op.constant([1, 1])
tf_c = constant_op.constant([[1, 1], [1, 1]])
tf_d = constant_op.constant([[1, 2], [1, 2]])
tf_e = constant_op.constant([1, 1, 1])
np_a = np.array([1, 1])
np_b = np.array([1, 1])
np_c = np.array([[1, 1], [1, 1]])
np_d = np.array([[1, 2], [1, 2]])
np_e = np.array([1, 1, 1])
ops.disable_tensor_equality()
# We don't do element-wise comparison
self.assertNotEqual(tf_a, tf_b)
self.assertNotEqual(tf_a, tf_c)
self.assertNotEqual(tf_a, tf_d)
ops.enable_tensor_equality()
# We do element-wise comparison but can't convert results array to bool
with self.assertRaises(ValueError):
bool(tf_a == tf_b)
self.assertAllEqual(tf_a == tf_b, [True, True])
with self.assertRaises(ValueError):
bool(tf_a == tf_c)
self.assertAllEqual(tf_a == tf_c, [[True, True], [True, True]])
with self.assertRaises(ValueError):
bool(tf_a == tf_d)
self.assertAllEqual(tf_a == tf_d, [[True, False], [True, False]])
if compat.forward_compatible(2019, 9, 25):
self.assertFalse(bool(tf_a == tf_e))
self.assertTrue(bool(tf_a != tf_e))
self.assertNotAllEqual(tf_a, tf_e)
else:
with self.assertRaises(errors.InvalidArgumentError):
bool(tf_a != tf_e)
with self.assertRaises(ValueError):
bool(np_a == np_b)
self.assertAllEqual(np_a == np_b, [True, True])
with self.assertRaises(ValueError):
bool(np_a == np_c)
self.assertAllEqual(np_a == np_c, [[True, True], [True, True]])
self.assertAllEqual(np_a == np_d, [[True, False], [True, False]])
self.assertFalse(bool(np_a == np_e))
self.assertTrue(bool(np_a != np_e))
self.assertNotAllEqual(np_a, np_e)
finally:
if default:
ops.enable_tensor_equality()
else:
ops.disable_tensor_equality()
def testEqualityBroadcast(self):
default = ops.Tensor._USE_EQUALITY
try:
tf_a = constant_op.constant([1, 1])
tf_b = constant_op.constant([1, 1])
tf_c = constant_op.constant([[1, 1], [1, 1]])
tf_d = constant_op.constant([[1, 2], [1, 2]])
tf_e = constant_op.constant([1, 1, 1])
np_a = np.array([1, 1])
np_b = np.array([1, 1])
np_c = np.array([[1, 1], [1, 1]])
np_d = np.array([[1, 2], [1, 2]])
np_e = np.array([1, 1, 1])
ops.disable_tensor_equality()
# We don't do element-wise comparison
self.assertNotEqual(tf_a, tf_b)
self.assertNotEqual(tf_a, tf_c)
self.assertNotEqual(tf_a, tf_d)
ops.enable_tensor_equality()
# We do element-wise comparison but can't convert results array to bool
with self.assertRaises(ValueError):
bool(tf_a == tf_b)
self.assertAllEqual(tf_a == tf_b, [True, True])
with self.assertRaises(ValueError):
bool(tf_a == tf_c)
self.assertAllEqual(tf_a == tf_c, [[True, True], [True, True]])
with self.assertRaises(ValueError):
bool(tf_a == tf_d)
self.assertAllEqual(tf_a == tf_d, [[True, False], [True, False]])
# TODO(b/207402791): re-enable once incompatible shapes supported by XLA.
if not test_util.is_xla_enabled():
self.assertFalse(bool(tf_a == tf_e))
self.assertTrue(bool(tf_a != tf_e))
self.assertNotAllEqual(tf_a, tf_e)
with self.assertRaises(ValueError):
bool(np_a == np_b)
self.assertAllEqual(np_a == np_b, [True, True])
with self.assertRaises(ValueError):
bool(np_a == np_c)
self.assertAllEqual(np_a == np_c, [[True, True], [True, True]])
self.assertAllEqual(np_a == np_d, [[True, False], [True, False]])
self.assertFalse(bool(np_a == np_e))
self.assertTrue(bool(np_a != np_e))
self.assertNotAllEqual(np_a, np_e)
finally:
if default:
ops.enable_tensor_equality()
else:
ops.disable_tensor_equality()
def disable_v2_behavior():
"""Disables TensorFlow 2.x behaviors.
This function can be called at the beginning of the program
(before `Tensors`, `Graphs` or other structures have been created,
and before devices have been initialized. It switches all global
behaviors that are different between TensorFlow 1.x and 2.x to
behave as intended for 1.x.
User can call this function to disable 2.x behavior during complex
migrations.
"""
# _v2_behavior_usage_gauge.get_cell("disable").set(True)
tf2.disable()
ops.disable_eager_execution()
tensor_shape.disable_v2_tensorshape() # Also switched by tf2
# FIXME[bug]:
# Warning: disable_resource_variables (from
# tensorflow.python.ops.variable_scope) is deprecated and will be
# removed in a future version.
# Instructions for updating:
# non-resource variables are not supported in the long term
#
# The function tf.compat.v1.disable_resource_variables() is
# depreciated instead you can mention use_resource=False in
# tf.get_variable() which will be forced to true when eager excecution
# is enabled by default in Tensorflow 2.x.
# variable_scope.disable_resource_variables()
ops.disable_tensor_equality()
# Disables TensorArrayV2 and control flow V2.
control_flow_v2_toggles.disable_control_flow_v2()
# Make sure internal uses of tf.data symbols map to V1 versions.
dataset_ops.Dataset = dataset_ops.DatasetV1
readers.FixedLengthRecordDataset = readers.FixedLengthRecordDatasetV1
readers.TFRecordDataset = readers.TFRecordDatasetV1
readers.TextLineDataset = readers.TextLineDatasetV1
counter.Counter = counter.CounterV1
interleave_ops.choose_from_datasets = \
interleave_ops.choose_from_datasets_v1
interleave_ops.sample_from_datasets = \
interleave_ops.sample_from_datasets_v1
random_ops.RandomDataset = random_ops.RandomDatasetV1
exp_readers.CsvDataset = exp_readers.CsvDatasetV1
exp_readers.SqlDataset = exp_readers.SqlDatasetV1
exp_readers.make_batched_features_dataset = (
exp_readers.make_batched_features_dataset_v1)
exp_readers.make_csv_dataset = exp_readers.make_csv_dataset_v1
def testEqualityBroadcast(self):
default = ops.Tensor._USE_EQUALITY
try:
tf_a = constant_op.constant([1, 1])
tf_b = constant_op.constant([1, 1])
tf_c = constant_op.constant([[1, 1], [1, 1]])
tf_d = constant_op.constant([[1, 2], [1, 2]])
tf_e = constant_op.constant([1, 1, 1])
np_a = np.array([1, 1])
np_b = np.array([1, 1])
np_c = np.array([[1, 1], [1, 1]])
np_d = np.array([[1, 2], [1, 2]])
np_e = np.array([1, 1, 1])
ops.disable_tensor_equality()
# We don't do element-wise comparison
self.assertNotEqual(tf_a, tf_b)
self.assertNotEqual(tf_a, tf_c)
self.assertNotEqual(tf_a, tf_d)
ops.enable_tensor_equality()
# We do element-wise comparison but can't convert results array to bool
with self.assertRaises(ValueError):
bool(tf_a == tf_b)
self.assertAllEqual(tf_a == tf_b, [True, True])
with self.assertRaises(ValueError):
bool(tf_a == tf_c)
self.assertAllEqual(tf_a == tf_c, [[True, True], [True, True]])
with self.assertRaises(ValueError):
bool(tf_a == tf_d)
self.assertAllEqual(tf_a == tf_d, [[True, False], [True, False]])
# TODO(b/120678848): If shapes do not match we should instead return False
with self.assertRaises(errors.InvalidArgumentError):
bool(tf_a != tf_e)
with self.assertRaises(ValueError):
bool(np_a == np_b)
self.assertAllEqual(np_a == np_b, [True, True])
with self.assertRaises(ValueError):
bool(np_a == np_c)
self.assertAllEqual(np_a == np_c, [[True, True], [True, True]])
self.assertAllEqual(np_a == np_d, [[True, False], [True, False]])
bool(np_a != np_e)
finally:
if default:
ops.enable_tensor_equality()
else:
ops.disable_tensor_equality()
def test_deprecated_arg_values_when_value_is_none(self, mock_warning):
@deprecation.deprecated_arg_values("2016-07-04",
"This is how you update...",
warn_once=True,
arg0=None)
def _fn(arg0): # pylint: disable=unused-argument
pass
ops.enable_tensor_equality()
initial_count = mock_warning.call_count
# Check that we avoid error from explicit `var == None` check.
_fn(arg0=variables.Variable(0))
self.assertEqual(initial_count, mock_warning.call_count)
_fn(arg0=None)
self.assertEqual(initial_count + 1, mock_warning.call_count)
ops.disable_tensor_equality()
def disable_v2_behavior():
"""Disables TensorFlow 2.x behaviors.
This function can be called at the beginning of the program (before `Tensors`,
`Graphs` or other structures have been created, and before devices have been
initialized. It switches all global behaviors that are different between
TensorFlow 1.x and 2.x to behave as intended for 1.x.
User can call this function to disable 2.x behavior during complex migrations.
"""
tf2.disable()
ops.disable_eager_execution()
tensor_shape.disable_v2_tensorshape() # Also switched by tf2
variable_scope.disable_resource_variables()
ops.disable_tensor_equality()
# Disables TensorArrayV2 and control flow V2.
control_flow_v2_toggles.disable_control_flow_v2()
def disable_v2_behavior():
"""Disables TensorFlow 2.x behaviors.
This function can be called at the beginning of the program (before `Tensors`,
`Graphs` or other structures have been created, and before devices have been
initialized. It switches all global behaviors that are different between
TensorFlow 1.x and 2.x to behave as intended for 1.x.
User can call this function to disable 2.x behavior during complex migrations.
@compatibility(TF2)
Using this function indicates that your software is not compatible
with eager execution and `tf.function` in TF2.
To migrate to TF2, rewrite your code to be compatible with eager execution.
Please refer to the [migration guide]
(https://www.tensorflow.org/guide/migrate) for additional resource on the
topic.
@end_compatibility
"""
_v2_behavior_usage_gauge.get_cell("disable").set(True)
tf2.disable()
ops.disable_eager_execution()
tensor_shape.disable_v2_tensorshape() # Also switched by tf2
variable_scope.disable_resource_variables()
ops.disable_tensor_equality()
# Disables TensorArrayV2 and control flow V2.
control_flow_v2_toggles.disable_control_flow_v2()
# Make sure internal uses of tf.data symbols map to V1 versions.
dataset_ops.Dataset = dataset_ops.DatasetV1
readers.FixedLengthRecordDataset = readers.FixedLengthRecordDatasetV1
readers.TFRecordDataset = readers.TFRecordDatasetV1
readers.TextLineDataset = readers.TextLineDatasetV1
counter.Counter = counter.CounterV1
interleave_ops.choose_from_datasets = interleave_ops.choose_from_datasets_v1
interleave_ops.sample_from_datasets = interleave_ops.sample_from_datasets_v1
random_ops.RandomDataset = random_ops.RandomDatasetV1
exp_readers.CsvDataset = exp_readers.CsvDatasetV1
exp_readers.SqlDataset = exp_readers.SqlDatasetV1
exp_readers.make_batched_features_dataset = (
exp_readers.make_batched_features_dataset_v1)
exp_readers.make_csv_dataset = exp_readers.make_csv_dataset_v1
def testV2BehaviorLogging(self):
with self.assertLogs(level='INFO') as logs:
try:
ops.enable_eager_execution()
# Ignore this exception to test log output successfully
except ValueError as e:
if 'must be called at program startup' not in str(e):
raise e
self.assertIn('Enabling eager execution', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
ops.disable_eager_execution()
self.assertIn('Disabling eager execution', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
tensor_shape.enable_v2_tensorshape()
self.assertIn('Enabling v2 tensorshape', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
tensor_shape.disable_v2_tensorshape()
self.assertIn('Disabling v2 tensorshape', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
variable_scope.enable_resource_variables()
self.assertIn('Enabling resource variables', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
variable_scope.disable_resource_variables()
self.assertIn('Disabling resource variables', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
ops.enable_tensor_equality()
self.assertIn('Enabling tensor equality', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
ops.disable_tensor_equality()
self.assertIn('Disabling tensor equality', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
control_flow_v2_toggles.enable_control_flow_v2()
self.assertIn('Enabling control flow v2', ''.join(logs.output))
with self.assertLogs(level='INFO') as logs:
control_flow_v2_toggles.disable_control_flow_v2()
self.assertIn('Disabling control flow v2', ''.join(logs.output))
def testEquality(self):
default = ops.Tensor._USE_EQUALITY
try:
def _v1_check(a, b):
self.assertEqual(a, a)
self.assertIs(a, a)
self.assertNotEqual(a, 1.0)
self.assertIsNot(a, 1.0)
self.assertNotEqual(a, b)
self.assertIsNot(a, b)
def _v2_check(a, b):
self.assertEqual(a, a)
self.assertIs(a, a)
self.assertEqual(a, 1.0)
self.assertIsNot(a, 1.0)
self.assertEqual(a, b)
self.assertIsNot(a, b)
constant_a = constant_op.constant(1.0)
constant_b = constant_op.constant(1.0)
ops.disable_tensor_equality()
self._test_hashable(constant_a, constant_b, True)
_v1_check(constant_a, constant_b)
ops.enable_tensor_equality()
_v2_check(constant_a, constant_b)
self._test_hashable(constant_a, constant_b, False)
variable_a = variables.Variable(1.0)
variable_b = variables.Variable(1.0)
ops.disable_tensor_equality()
_v1_check(variable_a, variable_b)
self._test_hashable(variable_a, variable_b, True)
ops.enable_tensor_equality()
_v2_check(variable_a, variable_b)
self._test_hashable(variable_a, variable_b, False)
# We only test numpy behaviour in v2 mode since we'd like to match that.
numpy_a = np.array(1.0)
numpy_b = np.array(1.0)
_v2_check(numpy_a, numpy_b)
self._test_hashable(numpy_a, numpy_b, False)
finally:
if default:
ops.enable_tensor_equality()
else:
ops.disable_tensor_equality()
def testEqualityNan(self):
default = ops.Tensor._USE_EQUALITY
try:
def _v1_check(a, b):
self.assertEqual(a, a)
self.assertIs(a, a)
self.assertNotEqual(a, float('nan'))
self.assertIsNot(a, float('nan'))
self.assertNotEqual(a, b)
self.assertIsNot(a, b)
def _v2_check(a, b):
self.assertNotEqual(a, a)
self.assertIs(a, a)
self.assertNotEqual(a, float('nan'))
self.assertIsNot(a, float('nan'))
self.assertNotEqual(a, b)
self.assertIsNot(a, b)
constant_a = constant_op.constant(float('nan'))
constant_b = constant_op.constant(float('nan'))
ops.disable_tensor_equality()
self._test_hashable(constant_a, constant_b, True)
_v1_check(constant_a, constant_b)
ops.enable_tensor_equality()
_v2_check(constant_a, constant_b)
self._test_hashable(constant_a, constant_b, False)
variable_a = variables.Variable(float('nan'))
variable_b = variables.Variable(float('nan'))
ops.disable_tensor_equality()
_v1_check(variable_a, variable_b)
self._test_hashable(variable_a, variable_b, True)
ops.enable_tensor_equality()
_v2_check(variable_a, variable_b)
self._test_hashable(variable_a, variable_b, False)
numpy_a = np.array(float('nan'))
numpy_b = np.array(float('nan'))
_v2_check(numpy_a, numpy_b)
self._test_hashable(numpy_a, numpy_b, False)
finally:
if default:
ops.enable_tensor_equality()
else:
ops.disable_tensor_equality()
def test_binary_cwise_ops(self):
# Enable tensor equality to test `equal` and `not_equal` ops below.
default_equality = framework_ops.Tensor._USE_EQUALITY
framework_ops.enable_tensor_equality()
try:
logical_ops = [
math_ops.logical_and, math_ops.logical_or, math_ops.logical_xor
]
# Wrapper functions restricting the range of inputs of zeta and polygamma.
def safe_polygamma(x, y):
return math_ops.polygamma(
math_ops.round(clip_ops.clip_by_value(y, 1, 10)), x * x + 1)
def safe_zeta(x, y):
return math_ops.zeta(x * x + 1, y * y)
float_ops = [
math_ops.add,
math_ops.add_v2,
math_ops.atan2,
math_ops.complex,
math_ops.div,
math_ops.divide,
math_ops.div_no_nan,
math_ops.equal,
lambda x, y: framework_ops.convert_to_tensor(x == y),
lambda x, y: framework_ops.convert_to_tensor(x != y),
math_ops.floor_mod,
math_ops.greater,
math_ops.greater_equal,
math_ops.igamma,
math_ops.igammac,
math_ops.igamma_grad_a,
math_ops.less,
math_ops.less_equal,
math_ops.maximum,
math_ops.minimum,
math_ops.mod,
math_ops.multiply,
math_ops.not_equal,
math_ops.pow,
math_ops.squared_difference,
math_ops.subtract,
math_ops.truncate_mod,
safe_polygamma,
safe_zeta,
]
# FloorDiv fails on XLA due floor's discontinuities exacerbating small
# division differences.
if not test_util.is_xla_enabled():
float_ops += [math_ops.floor_div]
for op in logical_ops + float_ops:
x = random_ops.random_uniform([7, 3, 5])
y = random_ops.random_uniform([3, 5])
if op in logical_ops:
x = x > 0
y = y > 0
output_dtypes = []
# pylint: disable=cell-var-from-loop
def loop_fn(i):
x1 = array_ops.gather(x, i)
y1 = array_ops.gather(y, i)
outputs = [op(x, y), op(x1, y), op(x, y1), op(x1, y1), op(x1, x1)]
del output_dtypes[:]
output_dtypes.extend(t.dtype for t in outputs)
return outputs
# pylint: enable=cell-var-from-loop
self._test_loop_fn(loop_fn, 3)
finally:
if not default_equality:
framework_ops.disable_tensor_equality()
def testEqualityCompare(self):
default = ops.Tensor._USE_EQUALITY
try:
tf_a = constant_op.constant([1, 2])
tf_b = constant_op.constant([1, 2])
tf_c = constant_op.constant([1, 1])
np_a = np.array([1, 2])
np_b = np.array([1, 2])
np_c = np.array([1, 1])
ops.disable_tensor_equality()
# We don't do element-wise comparison
self.assertNotEqual(tf_a, tf_b)
self.assertNotEqual(tf_a, tf_c)
# We can compare list of tensors
self.assertEqual([tf_a, tf_b], [tf_a, tf_b])
self.assertNotEqual([tf_a, tf_b], [tf_b, tf_b])
# We can compare existence in a list
self.assertIn(tf_a, [tf_a, tf_b])
self.assertIn(tf_a, [tf_b, tf_a])
self.assertNotIn(tf_a, [tf_b, tf_c])
ops.enable_tensor_equality()
# We do element-wise comparison but can't convert results array to bool
with self.assertRaises(ValueError):
bool(tf_a == tf_b)
self.assertAllEqual(tf_a == tf_b, [True, True])
with self.assertRaises(ValueError):
bool(tf_a == tf_c)
self.assertAllEqual(tf_a == tf_c, [True, False])
self.assertNotAllEqual(tf_a, tf_c)
with self.assertRaises(ValueError):
bool(np_a == np_b)
self.assertAllEqual(np_a == np_b, [True, True])
with self.assertRaises(ValueError):
bool(np_a == np_c)
self.assertAllEqual(np_a == np_c, [True, False])
self.assertNotAllEqual(np_a, np_c)
# Warning even though we technically shouldn't be able to compare here,
# since the id is the same both TF & numpy will handle lists with the same
# value without raising an error
self.assertEqual([tf_a, tf_b], [tf_a, tf_b])
with self.assertRaises(ValueError):
bool([tf_a, tf_b] == [tf_b, tf_b])
self.assertEqual([np_a, np_b], [np_a, np_b])
with self.assertRaises(ValueError):
bool([np_a, np_b] == [np_b, np_b])
# Similar to lists we shouldn't be able to do a `in` check such as
# `if a in [a,b]`. However if `a` is the first element, it works due to
# short circuiting
self.assertIn(tf_a, [tf_a, tf_b])
with self.assertRaises(ValueError):
bool(tf_a in [tf_b, tf_a])
with self.assertRaises(ValueError):
bool(tf_a in [tf_b, tf_c])
self.assertIn(np_a, [np_a, np_b])
with self.assertRaises(ValueError):
bool(np_a in [np_b, np_a])
with self.assertRaises(ValueError):
bool(np_a in [np_b, np_c])
# rank 0
self.assertAllEqual(
constant_op.constant(1) == constant_op.constant(1), True)
self.assertAllEqual(
constant_op.constant(1) == constant_op.constant(2), False)
self.assertAllEqual(np.array(1) == np.array(1), True)
self.assertAllEqual(np.array(1) == np.array(2), False)
finally:
if default:
ops.enable_tensor_equality()
else:
ops.disable_tensor_equality()
from tensorflow_probability import distributions as tfd
from unification import var, unify
from kanren import run, eq, lall
from kanren.graph import walko
from kanren.assoccomm import eq_comm, commutative
from symbolic_pymc.meta import enable_lvar_defaults
from symbolic_pymc.tensorflow.meta import mt
from symbolic_pymc.tensorflow.graph import normalize_tf_graph
from tests.tensorflow import run_in_graph_mode
from tests.tensorflow.utils import mt_normal_log_prob
disable_tensor_equality()
@run_in_graph_mode
def test_walko():
with enable_lvar_defaults("names"):
add_1_mt = mt(1) + mt(2)
def walk_rel(x, y):
return lall(eq(x, mt(1)), eq(y, mt(3)))
q = var()
(res, ) = run(1, q, walko(walk_rel, add_1_mt, q))
# The easiest way to check whether or not two arbitrary TF meta graphs are
# (structurally) equivalent is to confirm that they unify. This avoids
