def test_mnist_training_round_trip(self):
it = canonical_form_utils.get_iterative_process_for_canonical_form(
test_utils.get_mnist_training_example())
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
new_it = canonical_form_utils.get_iterative_process_for_canonical_form(
cf)
state1 = it.initialize()
state2 = new_it.initialize()
self.assertEqual(str(state1), str(state2))
dummy_x = np.array([[0.5] * 784], dtype=np.float32)
dummy_y = np.array([1], dtype=np.int32)
client_data = [
collections.OrderedDict([('x', dummy_x), ('y', dummy_y)])
]
round_1 = it.next(state1, [client_data])
state = round_1[0]
metrics = round_1[1]
alt_round_1 = new_it.next(state2, [client_data])
alt_state = alt_round_1[0]
alt_metrics = alt_round_1[1]
self.assertEqual(str(round_1), str(alt_round_1))
self.assertTrue(
np.array_equal(state.model.weights, state.model.weights))
self.assertTrue(np.array_equal(state.model.bias, alt_state.model.bias))
self.assertTrue(np.array_equal(state.num_rounds, alt_state.num_rounds))
self.assertTrue(
np.array_equal(metrics.num_rounds, alt_metrics.num_rounds))
self.assertTrue(
np.array_equal(metrics.num_examples, alt_metrics.num_examples))
self.assertTrue(np.array_equal(metrics.loss, alt_metrics.loss))
def test_temperature_example_round_trip(self):
it = canonical_form_utils.get_iterative_process_for_canonical_form(
test_utils.get_temperature_sensor_example())
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
new_it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
state = new_it.initialize()
self.assertLen(state, 1)
self.assertAllEqual(anonymous_tuple.name_list(state), ['num_rounds'])
self.assertEqual(state[0], 0)
state, metrics, stats = new_it.next(state, [[28.0], [30.0, 33.0, 29.0]])
self.assertLen(state, 1)
self.assertAllEqual(anonymous_tuple.name_list(state), ['num_rounds'])
self.assertEqual(state[0], 1)
self.assertLen(metrics, 1)
self.assertAllEqual(
anonymous_tuple.name_list(metrics), ['ratio_over_threshold'])
self.assertEqual(metrics[0], 0.5)
self.assertCountEqual([self.evaluate(x.num_readings) for x in stats],
[1, 3])
state, metrics, stats = new_it.next(state, [[33.0], [34.0], [35.0], [36.0]])
self.assertAllEqual(state, (2,))
self.assertAllClose(metrics, {'ratio_over_threshold': 0.75})
self.assertCountEqual([x.num_readings for x in stats], [1, 1, 1, 1])
self.assertEqual(
tree_analysis.count_tensorflow_variables_under(
test_utils.computation_to_building_block(it.next)),
tree_analysis.count_tensorflow_variables_under(
test_utils.computation_to_building_block(new_it.next)))
def test_mnist_training_round_trip(self):
it = canonical_form_utils.get_iterative_process_for_canonical_form(
test_utils.get_mnist_training_example())
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
new_it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
state1 = it.initialize()
state2 = new_it.initialize()
self.assertEqual(str(state1), str(state2))
dummy_x = np.array([[0.5] * 784], dtype=np.float32)
dummy_y = np.array([1], dtype=np.int32)
client_data = [collections.OrderedDict(x=dummy_x, y=dummy_y)]
round_1 = it.next(state1, [client_data])
state = round_1[0]
metrics = round_1[1]
alt_round_1 = new_it.next(state2, [client_data])
alt_state = alt_round_1[0]
alt_metrics = alt_round_1[1]
self.assertAllEqual(
anonymous_tuple.name_list(state), anonymous_tuple.name_list(alt_state))
self.assertAllEqual(
anonymous_tuple.name_list(metrics),
anonymous_tuple.name_list(alt_metrics))
self.assertAllClose(state, alt_state)
self.assertAllClose(metrics, alt_metrics)
self.assertEqual(
tree_analysis.count_tensorflow_variables_under(
test_utils.computation_to_building_block(it.next)),
tree_analysis.count_tensorflow_variables_under(
test_utils.computation_to_building_block(new_it.next)))
def test_with_temperature_sensor_example(self):
cf = test_utils.get_temperature_sensor_example()
it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
state = it.initialize()
self.assertLen(state, 1)
self.assertAllEqual(anonymous_tuple.name_list(state), ['num_rounds'])
self.assertEqual(state[0], 0)
state, metrics, stats = it.next(state, [[28.0], [30.0, 33.0, 29.0]])
self.assertLen(state, 1)
self.assertAllEqual(anonymous_tuple.name_list(state), ['num_rounds'])
self.assertEqual(state[0], 1)
self.assertLen(metrics, 1)
self.assertAllEqual(anonymous_tuple.name_list(metrics),
['ratio_over_threshold'])
self.assertEqual(metrics[0], 0.5)
self.assertCountEqual([self.evaluate(x.num_readings) for x in stats],
[1, 3])
state, metrics, stats = it.next(state,
[[33.0], [34.0], [35.0], [36.0]])
self.assertAllEqual(state, (2, ))
self.assertAllClose(metrics, {'ratio_over_threshold': 0.75})
self.assertCountEqual([x.num_readings for x in stats], [1, 1, 1, 1])
def test_passes_function_and_compiled_computation_of_same_type(self):
init = canonical_form_utils.get_iterative_process_for_canonical_form(
mapreduce_test_utils.get_temperature_sensor_example()).initialize
compiled_computation = self.compiled_computation_for_initialize(init)
function = building_blocks.Reference('f',
compiled_computation.type_signature)
transformations.check_extraction_result(function, compiled_computation)
def test_raises_non_function_and_compiled_computation(self):
init = canonical_form_utils.get_iterative_process_for_canonical_form(
mapreduce_test_utils.get_temperature_sensor_example()).initialize
compiled_computation = self.compiled_computation_for_initialize(init)
integer_ref = building_blocks.Reference('x', tf.int32)
with self.assertRaisesRegex(transformations.CanonicalFormCompilationError,
'we have the non-functional type'):
transformations.check_extraction_result(integer_ref, compiled_computation)
def test_temperature_example_round_trip(self):
it = canonical_form_utils.get_iterative_process_for_canonical_form(
test_utils.get_temperature_sensor_example())
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
new_it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
state = new_it.initialize()
self.assertEqual(str(state), '<num_rounds=0>')
state, metrics, stats = new_it.next(state, [[28.0], [30.0, 33.0, 29.0]])
self.assertEqual(str(state), '<num_rounds=1>')
self.assertEqual(str(metrics), '<ratio_over_threshold=0.5>')
self.assertCountEqual([x.num_readings for x in stats], [1, 3])
state, metrics, stats = new_it.next(state, [[33.0], [34.0], [35.0], [36.0]])
self.assertEqual(str(state), '<num_rounds=2>')
self.assertEqual(str(metrics), '<ratio_over_threshold=0.75>')
self.assertCountEqual([x.num_readings for x in stats], [1, 1, 1, 1])
def test_raises_function_and_compiled_computation_of_different_type(self):
init = canonical_form_utils.get_iterative_process_for_canonical_form(
mapreduce_test_utils.get_temperature_sensor_example()).initialize
compiled_computation = self.compiled_computation_for_initialize(init)
function = building_blocks.Reference(
'f', computation_types.FunctionType(tf.int32, tf.int32))
with self.assertRaisesRegex(transformations.CanonicalFormCompilationError,
'incorrect TFF type'):
transformations.check_extraction_result(function, compiled_computation)
def test_already_reduced_case(self):
init = canonical_form_utils.get_iterative_process_for_canonical_form(
mapreduce_test_utils.get_temperature_sensor_example()).initialize
comp = mapreduce_test_utils.computation_to_building_block(init)
result = transformations.consolidate_and_extract_local_processing(comp)
self.assertIsInstance(result, building_blocks.CompiledComputation)
self.assertIsInstance(result.proto, computation_pb2.Computation)
self.assertEqual(result.proto.WhichOneof('computation'), 'tensorflow')
def test_next_computation_returning_tensor_fails_well(self):
cf = test_utils.get_temperature_sensor_example()
it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
init_result = it.initialize.type_signature.result
lam = building_blocks.Lambda('x', init_result,
building_blocks.Reference('x', init_result))
bad_it = iterative_process.IterativeProcess(
it.initialize,
computation_wrapper_instances.building_block_to_computation(lam))
with self.assertRaises(TypeError):
canonical_form_utils.get_canonical_form_for_iterative_process(bad_it)
def test_returns_canonical_form_from_tff_learning_structure(self):
it = test_utils.construct_example_training_comp()
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
new_it = canonical_form_utils.get_iterative_process_for_canonical_form(
cf)
self.assertIsInstance(cf, canonical_form.CanonicalForm)
self.assertEqual(it.initialize.type_signature,
new_it.initialize.type_signature)
# Notice next type_signatures need not be equal, since we may have appended
# an empty tuple as client side-channel outputs if none existed
self.assertEqual(it.next.type_signature.parameter,
new_it.next.type_signature.parameter)
self.assertEqual(it.next.type_signature.result[0],
new_it.next.type_signature.result[0])
self.assertEqual(it.next.type_signature.result[1],
new_it.next.type_signature.result[1])
state1 = it.initialize()
state2 = new_it.initialize()
sample_batch = collections.OrderedDict(x=np.array([[1., 1.]],
dtype=np.float32),
y=np.array([[0]],
dtype=np.int32))
client_data = [sample_batch]
round_1 = it.next(state1, [client_data])
state = round_1[0]
state_names = anonymous_tuple.name_list(state)
state_arrays = anonymous_tuple.flatten(state)
metrics = round_1[1]
metrics_names = [x[0] for x in anonymous_tuple.iter_elements(metrics)]
metrics_arrays = anonymous_tuple.flatten(metrics)
alt_round_1 = new_it.next(state2, [client_data])
alt_state = alt_round_1[0]
alt_state_names = anonymous_tuple.name_list(alt_state)
alt_state_arrays = anonymous_tuple.flatten(alt_state)
alt_metrics = alt_round_1[1]
alt_metrics_names = [
x[0] for x in anonymous_tuple.iter_elements(alt_metrics)
]
alt_metrics_arrays = anonymous_tuple.flatten(alt_metrics)
self.assertEmpty(state.delta_aggregate_state)
self.assertEmpty(state.model_broadcast_state)
self.assertAllEqual(state_names, alt_state_names)
self.assertAllEqual(metrics_names, alt_metrics_names)
self.assertAllClose(state_arrays, alt_state_arrays)
self.assertAllClose(metrics_arrays[:2], alt_metrics_arrays[:2])
# Final metric is execution time
self.assertAlmostEqual(metrics_arrays[2],
alt_metrics_arrays[2],
delta=1e-3)
def test_get_canonical_form_from_fl_api(self):
it = test_utils.construct_example_training_comp()
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
new_it = canonical_form_utils.get_iterative_process_for_canonical_form(
cf)
self.assertIsInstance(cf, canonical_form.CanonicalForm)
self.assertEqual(it.initialize.type_signature,
new_it.initialize.type_signature)
# Notice next type_signatures need not be equal, since we may have appended
# an empty tuple as client side-channel outputs if none existed
self.assertEqual(it.next.type_signature.parameter,
new_it.next.type_signature.parameter)
self.assertEqual(it.next.type_signature.result[0],
new_it.next.type_signature.result[0])
self.assertEqual(it.next.type_signature.result[1],
new_it.next.type_signature.result[1])
state1 = it.initialize()
state2 = new_it.initialize()
self.assertEqual(str(state1), str(state2))
sample_batch = collections.OrderedDict([
('x', np.array([[1., 1.]], dtype=np.float32)),
('y', np.array([[0]], dtype=np.int32))
])
client_data = [sample_batch]
round_1 = it.next(state1, [client_data])
state = round_1[0]
metrics = round_1[1]
alt_round_1 = new_it.next(state2, [client_data])
alt_state = alt_round_1[0]
alt_metrics = alt_round_1[1]
self.assertTrue(
np.array_equal(state.model.trainable[0],
alt_state.model.trainable[0]))
self.assertTrue(
np.array_equal(state.model.trainable[1],
alt_state.model.trainable[1]))
self.assertEqual(str(state.model.non_trainable),
str(alt_state.model.non_trainable))
self.assertEqual(state.optimizer_state[0],
alt_state.optimizer_state[0])
self.assertEmpty(state.delta_aggregate_state)
self.assertEmpty(alt_state.delta_aggregate_state)
self.assertEmpty(state.model_broadcast_state)
self.assertEmpty(alt_state.model_broadcast_state)
self.assertEqual(metrics.sparse_categorical_accuracy,
alt_metrics.sparse_categorical_accuracy)
self.assertEqual(metrics.loss, alt_metrics.loss)
def test_temperature_example_round_trip(self):
# NOTE: the roundtrip through CanonicalForm->IterProc->CanonicalForm seems
# to lose the python container annotations on the StructType.
it = canonical_form_utils.get_iterative_process_for_canonical_form(
test_utils.get_temperature_sensor_example())
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
new_it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
state = new_it.initialize()
self.assertEqual(state.num_rounds, 0)
state, metrics = new_it.next(state, [[28.0], [30.0, 33.0, 29.0]])
self.assertEqual(state.num_rounds, 1)
self.assertAllClose(metrics,
collections.OrderedDict(ratio_over_threshold=0.5))
state, metrics = new_it.next(state, [[33.0], [34.0], [35.0], [36.0]])
self.assertAllClose(metrics,
collections.OrderedDict(ratio_over_threshold=0.75))
self.assertEqual(
tree_analysis.count_tensorflow_variables_under(
test_utils.computation_to_building_block(it.next)),
tree_analysis.count_tensorflow_variables_under(
test_utils.computation_to_building_block(new_it.next)))
def test_with_temperature_sensor_example(self):
cf = test_utils.get_temperature_sensor_example()
it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
state = it.initialize()
self.assertAllEqual(state, collections.OrderedDict(num_rounds=0))
state, metrics = it.next(state, [[28.0], [30.0, 33.0, 29.0]])
self.assertAllEqual(state, collections.OrderedDict(num_rounds=1))
self.assertAllClose(metrics,
collections.OrderedDict(ratio_over_threshold=0.5))
state, metrics = it.next(state, [[33.0], [34.0], [35.0], [36.0]])
self.assertAllClose(metrics,
collections.OrderedDict(ratio_over_threshold=0.75))
def test_broadcast_dependent_on_aggregate_fails_well(self):
cf = test_utils.get_temperature_sensor_example()
it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
next_comp = test_utils.computation_to_building_block(it.next)
top_level_param = building_blocks.Reference(next_comp.parameter_name,
next_comp.parameter_type)
first_result = building_blocks.Call(next_comp, top_level_param)
middle_param = building_blocks.Tuple([
building_blocks.Selection(first_result, index=0),
building_blocks.Selection(top_level_param, index=1)
])
second_result = building_blocks.Call(next_comp, middle_param)
not_reducible = building_blocks.Lambda(next_comp.parameter_name,
next_comp.parameter_type,
second_result)
not_reducible_it = iterative_process.IterativeProcess(
it.initialize,
computation_wrapper_instances.building_block_to_computation(
not_reducible))
with self.assertRaisesRegex(ValueError, 'broadcast dependent on aggregate'):
canonical_form_utils.get_canonical_form_for_iterative_process(
not_reducible_it)
def test_constructs_canonical_form_from_mnist_training_example(self):
it = canonical_form_utils.get_iterative_process_for_canonical_form(
test_utils.get_mnist_training_example())
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
self.assertIsInstance(cf, canonical_form.CanonicalForm)
def test_get_canonical_form_for_iterative_process(self):
cf = test_utils.get_temperature_sensor_example()
it = canonical_form_utils.get_iterative_process_for_canonical_form(cf)
cf = canonical_form_utils.get_canonical_form_for_iterative_process(it)
self.assertIsInstance(cf, canonical_form.CanonicalForm)
