def test_client_delta_clip_norm(self):
base_optimizer = optimizers.sgd(learning_rate=1.0)
train_batch_hparams = client_datasets.ShuffleRepeatBatchHParams(
batch_size=2, num_epochs=1, seed=0)
grad_batch_hparams = client_datasets.PaddedBatchHParams(batch_size=2)
server_learning_rate = 0.2
algorithm = mime_lite.mime_lite(per_example_loss,
base_optimizer,
train_batch_hparams,
grad_batch_hparams,
server_learning_rate,
client_delta_clip_norm=0.5)
clients = [
(b'cid0',
client_datasets.ClientDataset({'x': jnp.array([0.2, 0.4, 0.6])}),
jax.random.PRNGKey(0)),
(b'cid1',
client_datasets.ClientDataset({'x': jnp.array([0.8, 0.1])}),
jax.random.PRNGKey(1)),
]
state = algorithm.init({'w': jnp.array(4.)})
state, client_diagnostics = algorithm.apply(state, clients)
npt.assert_allclose(state.params['w'], 3.904)
npt.assert_allclose(
client_diagnostics[b'cid0']['clipped_delta_l2_norm'], 0.5)
npt.assert_allclose(
client_diagnostics[b'cid1']['clipped_delta_l2_norm'], 0.45000005)
def test_mime(self):
base_optimizer = optimizers.sgd(learning_rate=1.0)
train_batch_hparams = client_datasets.ShuffleRepeatBatchHParams(
batch_size=2, num_epochs=1, seed=0)
grad_batch_hparams = client_datasets.PaddedBatchHParams(batch_size=2)
server_learning_rate = 0.2
algorithm = mime.mime(per_example_loss, base_optimizer, train_batch_hparams,
grad_batch_hparams, server_learning_rate)
with self.subTest('init'):
state = algorithm.init({'w': jnp.array(4.)})
npt.assert_equal(state.params, {'w': jnp.array(4.)})
self.assertLen(state.opt_state, 2)
with self.subTest('apply'):
clients = [
(b'cid0',
client_datasets.ClientDataset({'x': jnp.array([2., 4., 6.])}),
jax.random.PRNGKey(0)),
(b'cid1', client_datasets.ClientDataset({'x': jnp.array([8., 10.])}),
jax.random.PRNGKey(1)),
]
state, client_diagnostics = algorithm.apply(state, clients)
npt.assert_allclose(state.params['w'], 2.08)
npt.assert_allclose(client_diagnostics[b'cid0']['delta_l2_norm'], 12.)
npt.assert_allclose(client_diagnostics[b'cid1']['delta_l2_norm'], 6.)
def test_model_train_clients_evaluation_fn(self):
sampler = FakeClientSampler()
sampler.set_round_num(1)
clients = [sampler.sample()[0] for _ in range(4)]
eval_fn = federated_experiment.ModelTrainClientsEvaluationFn(
fake_model(), client_datasets.PaddedBatchHParams(batch_size=4))
state = FakeState()
npt.assert_equal(eval_fn(state, 1, clients), {'accuracy': np.array(0.25)})
npt.assert_equal(eval_fn(state, 100, clients), {'accuracy': np.array(0.25)})
def test_model_full_evaluation_fn(self):
sampler = FakeClientSampler()
sampler.set_round_num(1)
clients = [sampler.sample()[0] for _ in range(4)]
fd = in_memory_federated_data.InMemoryFederatedData(
dict((k, v.all_examples()) for k, v, _ in clients))
eval_fn = federated_experiment.ModelFullEvaluationFn(
fd, fake_model(), client_datasets.PaddedBatchHParams(batch_size=4))
state = FakeState()
npt.assert_equal(eval_fn(state, 1), {'accuracy': np.array(0.25)})
npt.assert_equal(eval_fn(state, 100), {'accuracy': np.array(0.25)})
def test_model_sample_clients_evaluation_fn(self):
eval_fn = federated_experiment.ModelSampleClientsEvaluationFn(
FakeClientSampler(), fake_model(),
client_datasets.PaddedBatchHParams(batch_size=4))
state = FakeState()
npt.assert_equal(eval_fn(state, 1), {'accuracy': np.array(1.)})
npt.assert_equal(eval_fn(state, 2), {'accuracy': np.array(0.)})
npt.assert_equal(eval_fn(state, 3), {'accuracy': np.array(0.)})
npt.assert_equal(eval_fn(state, 4), {'accuracy': np.array(0.)})
npt.assert_equal(eval_fn(state, 5), {'accuracy': np.array(0.)})
npt.assert_equal(eval_fn(state, 6), {'accuracy': np.array(1.)})
def test_get(self):
with self.subTest('default'):
self.assertEqual(
self.SHUFFLE_BATCH_REPEAT.get(),
client_datasets.ShuffleRepeatBatchHParams(batch_size=128))
self.assertEqual(
self.PADDED_BATCH.get(),
client_datasets.PaddedBatchHParams(batch_size=128))
self.assertEqual(self.BATCH.get(),
client_datasets.BatchHParams(batch_size=128))
with self.subTest('custom'):
with flagsaver.flagsaver(shuffle_batch_repeat_batch_size=12,
shuffle_batch_repeat_num_epochs=21,
padded_batch_size=34,
batch_size=56):
self.assertEqual(
self.SHUFFLE_BATCH_REPEAT.get(),
client_datasets.ShuffleRepeatBatchHParams(batch_size=12,
num_epochs=21))
self.assertEqual(
self.PADDED_BATCH.get(),
client_datasets.PaddedBatchHParams(batch_size=34))
self.assertEqual(self.BATCH.get(),
client_datasets.BatchHParams(batch_size=56))
def test_padded_batch(self):
d = client_datasets.ClientDataset(
{
'a': np.arange(5),
'b': np.arange(10).reshape([5, 2])
}, client_datasets.BatchPreprocessor([lambda x: {
**x, 'a': 2 * x['a']
}]))
with self.subTest('1 bucket, kwargs'):
view = d.padded_batch(batch_size=3)
# `view` should be repeatedly iterable.
for _ in range(2):
batches = list(view)
self.assertLen(batches, 2)
npt.assert_equal(
batches[0], {
'a': [0, 2, 4],
'b': [[0, 1], [2, 3], [4, 5]],
'__mask__': [True, True, True],
})
npt.assert_equal(
batches[1], {
'a': [6, 8, 0],
'b': [[6, 7], [8, 9], [0, 0]],
'__mask__': [True, True, False]
})
with self.subTest('2 buckets, kwargs override'):
view = d.padded_batch(
client_datasets.PaddedBatchHParams(batch_size=4),
num_batch_size_buckets=2)
# `view` should be repeatedly iterable.
for _ in range(2):
batches = list(view)
self.assertLen(batches, 2)
npt.assert_equal(
batches[0], {
'a': [0, 2, 4, 6],
'b': [[0, 1], [2, 3], [4, 5], [6, 7]],
'__mask__': [True, True, True, True],
})
npt.assert_equal(batches[1], {
'a': [8, 0],
'b': [[8, 9], [0, 0]],
'__mask__': [True, False]
})
def test_agnostic_federated_averaging(self):
algorithm = agnostic_fed_avg.agnostic_federated_averaging(
per_example_loss=per_example_loss,
client_optimizer=optimizers.sgd(learning_rate=1.0),
server_optimizer=optimizers.sgd(learning_rate=0.1),
client_batch_hparams=client_datasets.ShuffleRepeatBatchHParams(
batch_size=3, num_epochs=1, seed=0),
domain_batch_hparams=client_datasets.PaddedBatchHParams(
batch_size=3),
init_domain_weights=[0.1, 0.2, 0.3, 0.4],
domain_learning_rate=0.01,
domain_algorithm='eg',
domain_window_size=2,
init_domain_window=[1., 2., 3., 4.])
with self.subTest('init'):
state = algorithm.init({'w': jnp.array(4.)})
npt.assert_equal(state.params, {'w': jnp.array(4.)})
self.assertLen(state.opt_state, 2)
npt.assert_allclose(state.domain_weights, [0.1, 0.2, 0.3, 0.4])
npt.assert_allclose(state.domain_window,
[[1., 2., 3., 4.], [1., 2., 3., 4.]])
with self.subTest('apply'):
clients = [
(b'cid0',
client_datasets.ClientDataset({
'x':
jnp.array([1., 2., 4., 3., 6., 1.]),
'domain_id':
jnp.array([1, 0, 0, 0, 2, 2])
}), jax.random.PRNGKey(0)),
(b'cid1',
client_datasets.ClientDataset({
'x':
jnp.array([8., 10., 5.]),
'domain_id':
jnp.array([1, 3, 1])
}), jax.random.PRNGKey(1)),
]
next_state, client_diagnostics = algorithm.apply(state, clients)
npt.assert_allclose(next_state.params['w'], 3.5555556)
npt.assert_allclose(
next_state.domain_weights,
[0.08702461, 0.18604803, 0.2663479, 0.46057943])
npt.assert_allclose(next_state.domain_window,
[[1., 2., 3., 4.], [3., 3., 2., 1.]])
npt.assert_allclose(client_diagnostics[b'cid0']['delta_l2_norm'],
2.8333335)
npt.assert_allclose(client_diagnostics[b'cid1']['delta_l2_norm'],
7.666667)
with self.subTest('invalid init_domain_weights'):
with self.assertRaisesRegex(
ValueError,
'init_domain_weights must sum to approximately 1.'):
agnostic_fed_avg.agnostic_federated_averaging(
per_example_loss=per_example_loss,
client_optimizer=optimizers.sgd(learning_rate=1.0),
server_optimizer=optimizers.sgd(learning_rate=1.0),
client_batch_hparams=client_datasets.
ShuffleRepeatBatchHParams(batch_size=3),
domain_batch_hparams=client_datasets.PaddedBatchHParams(
batch_size=3),
init_domain_weights=[50., 0., 0., 0.],
domain_learning_rate=0.5)
with self.subTest('unequal lengths'):
with self.assertRaisesRegex(
ValueError,
'init_domain_weights and init_domain_window must be equal lengths.'
):
agnostic_fed_avg.agnostic_federated_averaging(
per_example_loss=per_example_loss,
client_optimizer=optimizers.sgd(learning_rate=1.0),
server_optimizer=optimizers.sgd(learning_rate=1.0),
client_batch_hparams=client_datasets.
ShuffleRepeatBatchHParams(batch_size=3),
domain_batch_hparams=client_datasets.PaddedBatchHParams(
batch_size=3),
init_domain_weights=[0.1, 0.2, 0.3, 0.4],
domain_learning_rate=0.5,
init_domain_window=[1, 2])
def test_hyp_cluster(self):
functions_called = set()
def per_example_loss(params, batch, rng):
self.assertIsNotNone(rng)
functions_called.add('per_example_loss')
return jnp.square(params - batch['x'])
def regularizer(params):
del params
functions_called.add('regularizer')
return 0
client_optimizer = optimizers.sgd(0.5)
server_optimizer = optimizers.sgd(0.25)
maximization_batch_hparams = client_datasets.PaddedBatchHParams(
batch_size=2)
expectation_batch_hparams = client_datasets.ShuffleRepeatBatchHParams(
batch_size=1, num_epochs=5)
algorithm = hyp_cluster.hyp_cluster(
per_example_loss=per_example_loss,
client_optimizer=client_optimizer,
server_optimizer=server_optimizer,
maximization_batch_hparams=maximization_batch_hparams,
expectation_batch_hparams=expectation_batch_hparams,
regularizer=regularizer)
init_state = algorithm.init([jnp.array(1.), jnp.array(-1.)])
# Nothing happens with empty data.
no_op_state, diagnostics = algorithm.apply(init_state, clients=[])
npt.assert_array_equal(init_state.cluster_params,
no_op_state.cluster_params)
self.assertEmpty(diagnostics)
# Some actual training. PRNGKeys are not actually used.
clients = [
(b'0', client_datasets.ClientDataset({'x': np.array([1.1])}),
jax.random.PRNGKey(0)),
(b'1', client_datasets.ClientDataset({'x': np.array([0.9, 0.9])}),
jax.random.PRNGKey(1)),
(b'2', client_datasets.ClientDataset({'x': np.array([-1.1])}),
jax.random.PRNGKey(2)),
(b'3',
client_datasets.ClientDataset({'x': np.array([-0.9, -0.9,
-0.9])}),
jax.random.PRNGKey(3)),
]
next_state, diagnostics = algorithm.apply(init_state, clients)
npt.assert_equal(
diagnostics, {
b'0': {
'cluster_id': 0
},
b'1': {
'cluster_id': 0
},
b'2': {
'cluster_id': 1
},
b'3': {
'cluster_id': 1
},
})
cluster_params = next_state.cluster_params
self.assertIsInstance(cluster_params, list)
self.assertLen(cluster_params, 2)
npt.assert_allclose(cluster_params[0], [1. - 0.25 * 0.1 / 3])
npt.assert_allclose(cluster_params[1], [-1. + 0.25 * 0.2 / 4])
self.assertCountEqual(functions_called,
['per_example_loss', 'regularizer'])
def test_hyp_cluster_evaluator(self):
functions_called = set()
def apply_for_eval(params, batch):
functions_called.add('apply_for_eval')
score = params * batch['x']
return jnp.stack([-score, score], axis=-1)
def apply_for_train(params, batch, rng):
functions_called.add('apply_for_train')
self.assertIsNotNone(rng)
return params * batch['x']
def train_loss(batch, out):
functions_called.add('train_loss')
return jnp.abs(batch['y'] * 2 - 1 - out)
def regularizer(params):
# Just to check regularizer is called.
del params
functions_called.add('regularizer')
return 0
evaluator = hyp_cluster.HypClusterEvaluator(
models.Model(init=None,
apply_for_eval=apply_for_eval,
apply_for_train=apply_for_train,
train_loss=train_loss,
eval_metrics={'accuracy': metrics.Accuracy()}),
regularizer)
cluster_params = [jnp.array(1.), jnp.array(-1.)]
train_clients = [
# Evaluated using cluster 0.
(b'0',
client_datasets.ClientDataset({
'x': np.array([3., 2., 1.]),
'y': np.array([1, 1, 0])
}), jax.random.PRNGKey(0)),
# Evaluated using cluster 1.
(b'1',
client_datasets.ClientDataset({
'x':
np.array([0.9, -0.9, 0.8, -0.8, -0.3]),
'y':
np.array([0, 1, 0, 1, 0])
}), jax.random.PRNGKey(1)),
]
# Test clients are generated from train_clients by swapping client ids and
# then flipping labels.
test_clients = [
# Evaluated using cluster 0.
(b'0',
client_datasets.ClientDataset({
'x':
np.array([0.9, -0.9, 0.8, -0.8, -0.3]),
'y':
np.array([1, 0, 1, 0, 1])
})),
# Evaluated using cluster 1.
(b'1',
client_datasets.ClientDataset({
'x': np.array([3., 2., 1.]),
'y': np.array([0, 0, 1])
})),
]
for batch_size in [1, 2, 4]:
with self.subTest(f'batch_size = {batch_size}'):
batch_hparams = client_datasets.PaddedBatchHParams(
batch_size=batch_size)
metric_values = dict(
evaluator.evaluate_clients(cluster_params=cluster_params,
train_clients=train_clients,
test_clients=test_clients,
batch_hparams=batch_hparams))
self.assertCountEqual(metric_values, [b'0', b'1'])
self.assertCountEqual(metric_values[b'0'], ['accuracy'])
npt.assert_allclose(metric_values[b'0']['accuracy'], 4 / 5)
self.assertCountEqual(metric_values[b'1'], ['accuracy'])
npt.assert_allclose(metric_values[b'1']['accuracy'], 2 / 3)
self.assertCountEqual(
functions_called,
['apply_for_train', 'train_loss', 'apply_for_eval', 'regularizer'])
def test_maximization_step(self):
# L1 distance from centers.
def per_example_loss(params, batch, rng):
# Randomly flip the center to test rng behavior.
sign = jax.random.bernoulli(rng) * 2 - 1
return jnp.abs(params * sign - batch['x'])
def regularizer(params):
return 0.01 * jnp.sum(jnp.abs(params))
evaluator = models.AverageLossEvaluator(per_example_loss, regularizer)
cluster_params = [jnp.array(0.), jnp.array(-1.), jnp.array(2.)]
# Batch size is chosen so that we run 1 or 2 batches.
batch_hparams = client_datasets.PaddedBatchHParams(batch_size=2)
# Special seeds:
# - No flip in first 2 steps for all 3 clusters: 0;
# - Flip all in first 2 steps for all 3 clusters: 16;
# - No flip then flip all for all 3 clusters: 68;
# - Flips only cluster 1 in first 2 steps: 106.
clients = [
# No flip in first 2 steps for all 3 clusters.
(b'near0', client_datasets.ClientDataset({'x': np.array([0.1])}),
jax.random.PRNGKey(0)),
# Flip all in first 2 steps for all 3 clusters.
(b'near-1',
client_datasets.ClientDataset({'x': np.array([0.9, 1.1, 1.3])}),
jax.random.PRNGKey(16)),
# No flip then flip all for all 3 clusters.
(b'near2',
client_datasets.ClientDataset({'x': np.array([1.9, 2.1, -2.1])}),
jax.random.PRNGKey(68)),
# Flips only cluster 1 in first 2 steps.
(b'near1',
client_datasets.ClientDataset({'x': np.array([0.9, 1.1, 1.3])}),
jax.random.PRNGKey(106)),
]
cluster_losses = hyp_cluster._cluster_losses(
evaluator=evaluator,
cluster_params=cluster_params,
clients=clients,
batch_hparams=batch_hparams)
self.assertCountEqual(cluster_losses,
[b'near0', b'near-1', b'near2', b'near1'])
npt.assert_allclose(cluster_losses[b'near0'],
np.array([0.1, 1.1 + 0.01, 1.9 + 0.02]))
npt.assert_allclose(cluster_losses[b'near-1'],
np.array([1.1, 0.5 / 3 + 0.01, 3.1 + 0.02]))
npt.assert_allclose(cluster_losses[b'near2'],
np.array([6.1 / 3, 9.1 / 3 + 0.01, 0.1 + 0.02]),
rtol=1e-6)
npt.assert_allclose(cluster_losses[b'near1'],
np.array([1.1, 0.5 / 3 + 0.01, 0.9 + 0.02]))
client_cluster_ids = hyp_cluster.maximization_step(
evaluator=evaluator,
cluster_params=cluster_params,
clients=clients,
batch_hparams=batch_hparams)
self.assertDictEqual(client_cluster_ids, {
b'near0': 0,
b'near-1': 1,
b'near2': 2,
b'near1': 1
})
def test_kmeans_init(self):
functions_called = set()
def init(rng):
functions_called.add('init')
return jax.random.uniform(rng)
def apply_for_train(params, batch, rng):
functions_called.add('apply_for_train')
self.assertIsNotNone(rng)
return params - batch['x']
def train_loss(batch, out):
functions_called.add('train_loss')
return jnp.square(out) + batch['bias']
def regularizer(params):
del params
functions_called.add('regularizer')
return 0
initializer = hyp_cluster.ModelKMeansInitializer(
models.Model(init=init,
apply_for_train=apply_for_train,
apply_for_eval=None,
train_loss=train_loss,
eval_metrics={}), optimizers.sgd(0.5), regularizer)
# Each client has 1 example, so it's very easy to reach minimal loss, at
# which point the loss entirely depends on bias.
clients = [
(b'0',
client_datasets.ClientDataset({
'x': np.array([1.01]),
'bias': np.array([-2.])
}), jax.random.PRNGKey(1)),
(b'1',
client_datasets.ClientDataset({
'x': np.array([3.02]),
'bias': np.array([-1.])
}), jax.random.PRNGKey(2)),
(b'2',
client_datasets.ClientDataset({
'x': np.array([3.03]),
'bias': np.array([1.])
}), jax.random.PRNGKey(3)),
(b'3',
client_datasets.ClientDataset({
'x': np.array([1.04]),
'bias': np.array([2.])
}), jax.random.PRNGKey(3)),
]
train_batch_hparams = client_datasets.ShuffleRepeatBatchHParams(
batch_size=1, num_epochs=5)
eval_batch_hparams = client_datasets.PaddedBatchHParams(batch_size=2)
# Using a rng that leads to b'0' being the initial center.
cluster_params = initializer.cluster_params(
num_clusters=3,
rng=jax.random.PRNGKey(0),
clients=clients,
train_batch_hparams=train_batch_hparams,
eval_batch_hparams=eval_batch_hparams)
self.assertIsInstance(cluster_params, list)
self.assertLen(cluster_params, 3)
npt.assert_allclose(cluster_params, [1.01, 3.03, 1.04])
self.assertCountEqual(
functions_called,
['init', 'apply_for_train', 'train_loss', 'regularizer'])
# Using a rng that leads to b'2' being the initial center.
cluster_params = initializer.cluster_params(
num_clusters=3,
rng=jax.random.PRNGKey(1),
clients=clients,
train_batch_hparams=train_batch_hparams,
eval_batch_hparams=eval_batch_hparams)
self.assertIsInstance(cluster_params, list)
self.assertLen(cluster_params, 3)
npt.assert_allclose(cluster_params, [3.03, 1.04, 1.04])
def get(self):
return client_datasets.PaddedBatchHParams(
batch_size=self._get_flag('batch_size'))
