def test_broadcast_process_from_model(self):
model_fn = model_examples.LinearRegression
broadcast_process = (
encoding_utils.build_encoded_broadcast_process_from_model(
model_fn, _test_encoder_fn('simple')))
self.assertIsInstance(broadcast_process,
measured_process.MeasuredProcess)
def test_iterative_process_with_encoding_process(self):
model_fn = model_examples.LinearRegression
gather_process = encoding_utils.build_encoded_mean_process_from_model(
model_fn, _test_encoder_fn('gather'))
broadcast_process = (
encoding_utils.build_encoded_broadcast_process_from_model(
model_fn, _test_encoder_fn('simple')))
iterative_process = optimizer_utils.build_model_delta_optimizer_process(
model_fn=model_fn,
model_to_client_delta_fn=DummyClientDeltaFn,
server_optimizer_fn=lambda: tf.keras.optimizers.SGD(learning_rate=1.0),
aggregation_process=gather_process,
broadcast_process=broadcast_process)
self._verify_iterative_process(iterative_process)
def test_federated_evaluation_quantized_aggressively(self):
# Set up a uniform quantization encoder as the broadcaster.
broadcaster = (
encoding_utils.build_encoded_broadcast_process_from_model(
TestModelQuant, _build_simple_quant_encoder(2)))
self.assert_types_equivalent(broadcaster.next.type_signature,
_build_expected_broadcaster_next_signature())
evaluate = federated_evaluation.build_federated_evaluation(
TestModelQuant, broadcast_process=broadcaster)
# Confirm that the type signature matches what is expected.
self.assert_types_identical(
evaluate.type_signature,
_build_expected_test_quant_model_eval_signature())
def _temp_dict(temps):
return {'temp': np.array(temps, dtype=np.float32)}
result = evaluate(
collections.OrderedDict(
trainable=[[5.0, 10.0, 5.0, 7.0]], non_trainable=[]), [
[
_temp_dict([1.0, 10.0, 2.0, 7.0]),
_temp_dict([6.0, 11.0, 5.0, 8.0])
],
[_temp_dict([9.0, 12.0, 13.0, 7.0])],
[
_temp_dict([1.0, 22.0, 23.0, 24.0]),
_temp_dict([5.0, 10.0, 5.0, 7.0])
],
])
# This very aggressive quantization should be so lossy that some of the
# data is changed during encoding so the number that are equal between
# the original and the final result should not be 8 as it is in the
# conservative quantization test above.
self.assertContainsSubset(result.keys(), ['eval', 'stat'])
self.assertContainsSubset(result['eval'].keys(), ['num_same'])
self.assertLess(result['eval']['num_same'], 8.0)
self.assertContainsSubset(result['stat'].keys(), ['num_examples'])
self.assertEqual(result['stat']['num_examples'], 20)
def test_federated_evaluation_quantized_conservatively(self):
# Set up a uniform quantization encoder as the broadcaster.
broadcaster = (
encoding_utils.build_encoded_broadcast_process_from_model(
TestModelQuant, _build_simple_quant_encoder(12)))
type_test_utils.assert_types_equivalent(
broadcaster.next.type_signature,
_build_expected_broadcaster_next_signature())
evaluate = federated_evaluation.build_federated_evaluation(
TestModelQuant, broadcast_process=broadcaster)
# Confirm that the type signature matches what is expected.
type_test_utils.assert_types_identical(
evaluate.type_signature,
_build_expected_test_quant_model_eval_signature())
def _temp_dict(temps):
return {'temp': np.array(temps, dtype=np.float32)}
result = evaluate(
collections.OrderedDict(trainable=[[5.0, 10.0, 5.0, 7.0]],
non_trainable=[]),
[
[
_temp_dict([1.0, 10.0, 2.0, 7.0]),
_temp_dict([6.0, 11.0, 5.0, 8.0])
],
[_temp_dict([9.0, 12.0, 13.0, 7.0])],
[
_temp_dict([1.0, 22.0, 23.0, 24.0]),
_temp_dict([5.0, 10.0, 5.0, 7.0])
],
])
# This conservative quantization should not be too lossy.
# When comparing the data examples to trainable, there are 8 times
# where the index and value match.
self.assertEqual(
result,
collections.OrderedDict(eval=collections.OrderedDict(
num_same=8.0)))
