def test_returns_comp_with_block_untransformed(self):
data = building_blocks.Data('a', tf.int32)
block = building_blocks.Block([('x', data), ('y', data)], data)
untransformed, modified_indicator = compiler_transformations.remove_duplicate_called_graphs(
block)
self.assertEqual(untransformed, block)
self.assertFalse(modified_indicator)
def test_returns_higher_level_lambda_untransformed(self):
lower_level_lambda = building_blocks.Lambda(
'x', tf.int32, building_blocks.Reference('x', tf.int32))
higher_level_lambda = building_blocks.Lambda('y', tf.int32,
lower_level_lambda)
untransformed, modified_indicator = compiler_transformations.remove_duplicate_called_graphs(
higher_level_lambda)
self.assertEqual(untransformed, higher_level_lambda)
self.assertFalse(modified_indicator)
def test_returns_tf_computation_with_functional_type(self):
param = building_blocks.Reference('x', [('a', tf.int32),
('b', tf.float32)])
sel = building_blocks.Selection(source=param, index=0)
tup = building_blocks.Tuple([sel, sel, sel])
lam = building_blocks.Lambda(param.name, param.type_signature, tup)
transformed, modified_indicator = compiler_transformations.remove_duplicate_called_graphs(
lam)
self.assertTrue(modified_indicator)
self.assertIsInstance(transformed, building_blocks.CompiledComputation)
self.assertEqual(transformed.type_signature, lam.type_signature)
def _count_ops_parameterized_by_layers(k):
inlined_tuple_with_k_layers = _construct_inlined_tuple(k)
tf_representing_block_with_k_layers, _ = compiler_transformations.remove_duplicate_called_graphs(
inlined_tuple_with_k_layers)
block_ops_with_k_layers = tree_analysis.count_tensorflow_ops_under(
tf_representing_block_with_k_layers)
parser_callable = transformations.TFParser()
naively_generated_tf_with_k_layers, _ = transformation_utils.transform_postorder(
inlined_tuple_with_k_layers, parser_callable)
naive_ops_with_k_layers = tree_analysis.count_tensorflow_ops_under(
naively_generated_tf_with_k_layers)
return block_ops_with_k_layers, naive_ops_with_k_layers
def test_returns_called_tf_computation_with_non_functional_type(self):
constant_tuple = building_block_factory.create_tensorflow_constant(
[tf.int32, tf.float32], 1)
sel = building_blocks.Selection(source=constant_tuple, index=0)
tup = building_blocks.Tuple([sel, sel, sel])
transformed, modified_indicator = compiler_transformations.remove_duplicate_called_graphs(
tup)
self.assertTrue(modified_indicator)
self.assertEqual(transformed.type_signature, tup.type_signature)
self.assertIsInstance(transformed, building_blocks.Call)
self.assertIsInstance(transformed.function,
building_blocks.CompiledComputation)
self.assertIsNone(transformed.argument)
def test_raises_non_unique_names(self):
data = building_blocks.Data('a', tf.int32)
block = building_blocks.Block([('x', data), ('x', data)], data)
with self.assertRaises(ValueError):
compiler_transformations.remove_duplicate_called_graphs(block)
def test_raises_bad_type(self):
with self.assertRaises(TypeError):
compiler_transformations.remove_duplicate_called_graphs(1)
