# Copyright 2021 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Mime implementation.
Mime: Mimicking Centralized Stochastic Algorithms in Federated Learning
Sai Praneeth Karimireddy, Martin Jaggi, Satyen Kale, Mehryar Mohri,
Sashank J. Reddi, Sebastian U. Stich, Ananda Theertha Suresh
https://arxiv.org/abs/2008.03606
"""
from typing import Any, Callable, Mapping, Optional, Sequence, Tuple
from fedjax.core import client_datasets
from fedjax.core import dataclasses
from fedjax.core import federated_algorithm
from fedjax.core import federated_data
from fedjax.core import for_each_client
from fedjax.core import models
from fedjax.core import optimizers
from fedjax.core import tree_util
from fedjax.core.typing import BatchExample
from fedjax.core.typing import Params
from fedjax.core.typing import PRNGKey
import jax
import jax.numpy as jnp
Grads = Params
def create_grads_for_each_client(grad_fn):
"""Builds for_each_client for gradient computation."""
def client_init(server_params, client_rng):
client_step_state = {
'params': server_params,
'rng': client_rng,
'num_sum': 0.,
'grads_sum': jax.tree_util.tree_map(jnp.zeros_like, server_params)
}
return client_step_state
def client_step(client_step_state, batch):
rng, use_rng = jax.random.split(client_step_state['rng'])
grads = grad_fn(client_step_state['params'], batch, use_rng)
num = jnp.sum(batch[client_datasets.EXAMPLE_MASK_KEY])
grads_sum = tree_util.tree_add(
tree_util.tree_weight(grads, num), client_step_state['grads_sum'])
next_client_step_state = {
'params': client_step_state['params'],
'rng': rng,
'num_sum': client_step_state['num_sum'] + num,
'grads_sum': grads_sum
}
return next_client_step_state
def client_final(server_params, client_step_state):
del server_params
client_output = (client_step_state['grads_sum'],
client_step_state['num_sum'])
return client_output
return for_each_client.for_each_client(client_init, client_step, client_final)
def create_train_for_each_client(grad_fn, base_optimizer):
"""Builds for_each_client for client training."""
def client_init(shared_input, client_rng):
client_step_state = {
'params': shared_input['params'],
'opt_state': shared_input['opt_state'],
'rng': client_rng,
'init_params': shared_input['params'],
'control_variate': shared_input['control_variate'],
}
return client_step_state
def client_step(client_step_state, batch):
rng, use_rng = jax.random.split(client_step_state['rng'])
client_control_variate = grad_fn(client_step_state['init_params'], batch,
use_rng)
grads = grad_fn(client_step_state['params'], batch, use_rng)
adjusted_grads = jax.tree_util.tree_map(
lambda g, cc, c: g - cc + c, grads, client_control_variate,
client_step_state['control_variate'])
_, params = base_optimizer.apply(adjusted_grads,
client_step_state['opt_state'],
client_step_state['params'])
next_client_step_state = {
'params': params,
'opt_state': client_step_state['opt_state'],
'rng': rng,
'init_params': client_step_state['init_params'],
'control_variate': client_step_state['control_variate'],
}
return next_client_step_state
def client_final(shared_input, client_step_state):
delta_params = jax.tree_util.tree_map(lambda a, b: a - b,
shared_input['params'],
client_step_state['params'])
return delta_params
return for_each_client.for_each_client(client_init, client_step, client_final)
[docs]@dataclasses.dataclass
class ServerState:
"""State of server passed between rounds.
Attributes:
params: A pytree representing the server model parameters.
opt_state: A pytree representing the base optimizer state.
"""
params: Params
opt_state: optimizers.OptState
[docs]def mime(
per_example_loss: Callable[[Params, BatchExample, PRNGKey], jnp.ndarray],
base_optimizer: optimizers.Optimizer,
client_batch_hparams: client_datasets.ShuffleRepeatBatchHParams,
grads_batch_hparams: client_datasets.PaddedBatchHParams,
server_learning_rate: float,
regularizer: Optional[Callable[[Params], jnp.ndarray]] = None
) -> federated_algorithm.FederatedAlgorithm:
"""Builds mime.
Args:
per_example_loss: A function from (params, batch_example, rng) to a vector
of loss values for each example in the batch. This is used in both the
server gradient computation and gradient descent training.
base_optimizer: Base optimizer to mimic.
client_batch_hparams: Hyperparameters for batching client dataset for train.
grads_batch_hparams: Hyperparameters for batching client dataset for server
gradient computation.
server_learning_rate: Server learning rate.
regularizer: Optional regularizer that only depends on params.
Returns:
FederatedAlgorithm
"""
grad_fn = models.grad(per_example_loss, regularizer)
grads_for_each_client = create_grads_for_each_client(grad_fn)
train_for_each_client = create_train_for_each_client(grad_fn, base_optimizer)
def init(params: Params) -> ServerState:
opt_state = base_optimizer.init(params)
return ServerState(params, opt_state)
def apply(
server_state: ServerState,
clients: Sequence[Tuple[federated_data.ClientId,
client_datasets.ClientDataset, PRNGKey]]
) -> Tuple[ServerState, Mapping[federated_data.ClientId, Any]]:
# Compute full-batch gradient at server params on train data.
grads_batch_clients = [(cid, cds.padded_batch(grads_batch_hparams), crng)
for cid, cds, crng in clients]
grads_sum_total, num_sum_total = tree_util.tree_sum(
(co for _, co in grads_for_each_client(server_state.params,
grads_batch_clients)))
server_grads = tree_util.tree_inverse_weight(grads_sum_total, num_sum_total)
# Control variant corrected training across clients.
client_diagnostics = {}
client_num_examples = {cid: len(cds) for cid, cds, _ in clients}
batch_clients = [(cid, cds.shuffle_repeat_batch(client_batch_hparams), crng)
for cid, cds, crng in clients]
shared_input = {
'params': server_state.params,
'opt_state': server_state.opt_state,
'control_variate': server_grads
}
# Running weighted mean of client updates.
delta_params_sum = tree_util.tree_zeros_like(server_state.params)
num_examples_sum = 0.
for client_id, delta_params in train_for_each_client(
shared_input, batch_clients):
num_examples = client_num_examples[client_id]
delta_params_sum = tree_util.tree_add(
delta_params_sum, tree_util.tree_weight(delta_params, num_examples))
num_examples_sum += num_examples
client_diagnostics[client_id] = {
'delta_l2_norm': tree_util.tree_l2_norm(delta_params)
}
mean_delta_params = tree_util.tree_inverse_weight(delta_params_sum,
num_examples_sum)
server_state = server_update(server_state, server_grads, mean_delta_params)
return server_state, client_diagnostics
def server_update(server_state, server_grads, mean_delta_params):
# Server params uses weighted average of client updates, scaled by the
# server_learning_rate.
params = jax.tree_util.tree_map(
lambda p, q: p - server_learning_rate * q, server_state.params,
mean_delta_params)
opt_state, _ = base_optimizer.apply(server_grads, server_state.opt_state,
server_state.params)
return ServerState(params, opt_state)
return federated_algorithm.FederatedAlgorithm(init, apply)