Module fathom.algorithms.fathom_fedavg

Expand source code
# Copyright 2022 FATHOM Authors
#
# 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.

from typing import Any, Callable, Mapping, Sequence, Tuple, Union, Optional

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 optimizers
from fedjax.core import tree_util
from fedjax.core.typing import BatchExample
from fedjax.core.typing import Params, PyTree
from fedjax.core.typing import PRNGKey
from fedjax.core.typing import OptState, BatchExample
import jax
import jax.numpy as jnp
import fathom

Grads = Params

def create_train_for_each_client(grad_fn: Params, client_optimizer: optimizers.Optimizer):
    """Builds client_init, client_step, client_final for for_each_client."""
    normalize_hypergrad = 1

    def client_init(shared_input, client_rng):
        opt_state = client_optimizer.init(shared_input['params'])
        opt_state.hyperparams['learning_rate'] = shared_input['eta_c'] # Already sigmoided
        client_step_state = {
            'params': shared_input['params'],
            'params0': shared_input['params'],
            'opt_state': opt_state,
            'rng': client_rng,
            'step_idx': 0,
            'min_hypergrad': tree_util.tree_l2_norm(shared_input['params']), # Large enough value
            }
        return client_step_state

    def client_step(client_step_state, batch):
        rng, use_rng, metrics_rng = jax.random.split(client_step_state['rng'], num=3)
        grad_opt = grad_fn(client_step_state['params'], batch, use_rng)
        opt_state, params = client_optimizer.apply(grad_opt, client_step_state['opt_state'], client_step_state['params'])
        delta_params = jax.tree_util.tree_multimap(jnp.subtract, client_step_state['params0'], params)
        grad_opt_norm = tree_util.tree_l2_norm(grad_opt)
        delta_params_norm = tree_util.tree_l2_norm(delta_params)
        hypergrad = jnp.where(grad_opt_norm == 0., jnp.array(0),
            jnp.where(delta_params_norm == 0, jnp.array(0),
                jnp.where(normalize_hypergrad == 0 ,
                    fathom.core.tree_util.tree_dot(grad_opt, delta_params),
                    fathom.core.tree_util.tree_dot(grad_opt, delta_params) / grad_opt_norm / delta_params_norm)
            )
        )
        min_hypergrad = jnp.where(hypergrad > client_step_state['min_hypergrad'],
            client_step_state['min_hypergrad'], hypergrad
        )
        next_client_step_state = {
            'params': params,
            'params0': client_step_state['params0'],
            'opt_state': opt_state,
            'rng': rng,
            'step_idx': client_step_state['step_idx'] + 1,
            'min_hypergrad': min_hypergrad,
        }
        return next_client_step_state

    def client_final(shared_input, client_step_state) -> Tuple[Params, jnp.ndarray]:
        delta_params = jax.tree_util.tree_multimap(jnp.subtract, shared_input['params'], client_step_state['params'])
        return delta_params, client_step_state['min_hypergrad'], client_step_state['step_idx']

    return for_each_client.for_each_client(client_init, client_step, client_final)


@dataclasses.dataclass
class HyperParams:
    eta_c: float                # Local learning rate or step size
    Ep: float                  # Number of local steps = ceil(local_samples / batch_size) * Ep, where Ep ~ num epochs worth of data.
    bs: float                   # Local batch size
    alpha: float                # Momentum for glob grad estimation
    eta_h: jnp.ndarray          # Hyper optimizer learning rates for Ep, eta_c, and bs, respectively
    hparam_ub: jnp.ndarray      # Upperbound vals for Ep, eta_c, and bs, respectively.


@dataclasses.dataclass
class HyperState:
    hyperparams: HyperParams 
    init_hparams: HyperParams
    opt_state: optimizers.OptState
    opt_param: jnp.ndarray
    hypergrad_glob: float
    hypergrad_local: float


@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 server optimizer state.
    """
    params: Params
    params_bak: Params
    opt_state: optimizers.OptState
    round_index: int
    hyper_state: HyperState
    grad_glob: Params


@jax.jit
def estimate_grad_glob(server_state: ServerState, mean_delta_params: Params) -> Params:
    grad_glob = tree_util.tree_weight(server_state.grad_glob, server_state.hyper_state.hyperparams.alpha)
    delta_params = tree_util.tree_weight(mean_delta_params, 1. - server_state.hyper_state.hyperparams.alpha)
    grad_glob = tree_util.tree_add(grad_glob, delta_params)
    # grad_glob = fathom.core.tree_util.tree_inverse_weight(
    #     grad_glob, 
    #     (1. - server_state.hyper_state.hyperparams.alpha ** server_state.round_index)
    # )
    return grad_glob


def federated_averaging(
        grad_fn: Callable[[Params, BatchExample, PRNGKey], Grads],
        client_optimizer: optimizers.Optimizer,
        server_optimizer: optimizers.Optimizer,
        hyper_optimizer: optimizers.Optimizer,
        client_batch_hparams: client_datasets.ShuffleRepeatBatchHParams,
        fathom_init_hparams: HyperParams,
) -> federated_algorithm.FederatedAlgorithm:
    """Builds federated averaging.

    Args:
        grad_fn: A function from (params, batch_example, rng) to gradients.
            This can be created with :func:`fedjax.core.model.model_grad`.
        client_optimizer: Optimizer for local client training.
        server_optimizer: Optimizer for server update.
        client_batch_hparams: Hyperparameters for batching client dataset for train.

    Returns:
        FederatedAlgorithm
    """
    train_for_each_client = create_train_for_each_client(grad_fn, client_optimizer)

    def server_reset(params: Params, init_hparams: HyperParams) -> ServerState:
        opt_state_server = server_optimizer.init(params)
        opt_param_hyper = jnp.log(jnp.array([init_hparams.Ep, init_hparams.eta_c, init_hparams.bs]))
        opt_state_hyper = hyper_optimizer.init(opt_param_hyper)
        hyper_state = HyperState(
            hyperparams = init_hparams,
            init_hparams = init_hparams,
            opt_state = opt_state_hyper,
            opt_param = opt_param_hyper,
            hypergrad_glob = 0.,
            hypergrad_local = 0.,
        )
        # Need to initialize round_index to 1 for bias comp
        return ServerState(
            params = params, 
            params_bak = params,
            opt_state = opt_state_server, 
            round_index = 1, 
            grad_glob = tree_util.tree_zeros_like(params),
            hyper_state = hyper_state,
        )        

    def init(params: Params) -> ServerState:
        return server_reset(params, fathom_init_hparams)

    def apply(
        server_state: ServerState,
        clients: Sequence[Tuple[
            federated_data.ClientId, 
            client_datasets.ClientDataset, 
            PRNGKey
        ]],
    ) -> Tuple[ServerState, Mapping[federated_data.ClientId, Any]]:
        client_num_examples = {cid: len(cds) for cid, cds, _ in clients}
        Ep: float = server_state.hyper_state.hyperparams.Ep
        bs: int = max(int(server_state.hyper_state.hyperparams.bs + 0.5), 1) if server_state.hyper_state.hyperparams.bs > 0 else -1
        eta_c: float = server_state.hyper_state.hyperparams.eta_c
        batch_clients = [(cid, cds.shuffle_repeat_batch(
            client_datasets.ShuffleRepeatBatchHParams(
                batch_size = bs if bs > 0 else len(cds), 
                num_steps = max(int(jnp.ceil(Ep * client_num_examples[cid] / bs)), 1),
                num_epochs = None, # This is required.  See ShuffleRepeatBatchView implementation in fedjax.core.client_datasets.py.
                drop_remainder = client_batch_hparams.drop_remainder,
                seed = client_batch_hparams.seed,
                skip_shuffle = client_batch_hparams.skip_shuffle,
            )
        ), crng) for cid, cds, crng in clients]
        shared_input = {'params': server_state.params, 'eta_c': eta_c, 'Ep': Ep, 'bs': bs}
        client_diagnostics = {}
        # Running weighted mean of client updates. We do this iteratively to avoid
        # loading all the client outputs into memory since they can be prohibitively
        # large depending on the model parameters size.
        delta_params_sum = tree_util.tree_zeros_like(server_state.params)
        hypergrad_local_sum = jnp.array(0.0)
        num_examples_sum = 0.
        for client_id, (delta_params, min_hypergrad, num_steps) in train_for_each_client(shared_input, batch_clients):
            # Server collecting stats before sending them to server_update for updating params and metrics.
            num_examples = client_num_examples[client_id]
            weighted_delta_params =  tree_util.tree_weight(delta_params, num_examples)
            delta_params_sum = tree_util.tree_add(delta_params_sum, weighted_delta_params)
            hypergrad_local_sum = hypergrad_local_sum + min_hypergrad * num_examples
            num_examples_sum += num_examples
            # We record the l2 norm of client updates as an example, but it is not
            # required for the algorithm.
            client_diagnostics[client_id] = {
                'delta_l2_norm': tree_util.tree_l2_norm(delta_params),
            }
        mean_delta_params = fathom.core.tree_util.tree_inverse_weight(delta_params_sum, num_examples_sum)
        mean_hypergrad_local = hypergrad_local_sum / num_examples_sum
        server_state = server_update(
            server_state = server_state, 
            mean_delta_params = mean_delta_params, 
            hypergrad_local = mean_hypergrad_local, 
        )
        return server_state, client_diagnostics

    def server_update(
        server_state: ServerState, 
        mean_delta_params: Params, 
        hypergrad_local: jnp.ndarray, 
    ) -> ServerState:
        opt_state_server, params = server_optimizer.apply(
            mean_delta_params, 
            server_state.opt_state, 
            server_state.params,
        )
        grad_glob: Params = estimate_grad_glob(server_state, mean_delta_params)
        hyper_state: HyperState = hyper_update(
            server_state = server_state, 
            params = params, 
            delta_params = mean_delta_params,
            hypergrad_local = hypergrad_local,
        )
        return ServerState(
            params = params,
            params_bak = server_state.params_bak, # Keep initial params rather than update
            opt_state = opt_state_server,
            round_index = server_state.round_index + 1,
            hyper_state = hyper_state,
            grad_glob = grad_glob,
        )

    def hyper_update(
        server_state: ServerState,
        params: Params,
        delta_params: Params,
        hypergrad_local: jnp.ndarray,
    ) -> HyperState:

        opt_param, opt_state = server_state.hyper_state.opt_param, server_state.hyper_state.opt_state
        # Do not use the most current grad_glob as the result will bias positive
        hypergrad_glob: float = fathom.core.tree_util.tree_dot(server_state.grad_glob, delta_params)
        # Normalizing hypergrad_global here
        # hypergrad_local already normalized from local calculations
        grad_glob_norm = tree_util.tree_l2_norm(server_state.grad_glob)
        delta_params_norm = tree_util.tree_l2_norm(delta_params)
        hypergrad_glob = jnp.where((grad_glob_norm > 0. and delta_params_norm > 0.),
            hypergrad_glob / grad_glob_norm / delta_params_norm,
            0
        )
        hypergrad = - jnp.array([
            hypergrad_glob + hypergrad_local,   # Ep
            hypergrad_glob,                     # eta_c
            -hypergrad_local,                   # bs
        ]) 
        # This is where individual learning rates are applied, assuming opt is SGD.
        # With any other opt, individual learning rates need to be set at opt instantiation.
        hypergrad = hypergrad * server_state.hyper_state.hyperparams.eta_h

        # EGN gradients are already normalized
        opt_state, opt_param = hyper_optimizer.apply(hypergrad, opt_state, opt_param)
        hparams_vals = jnp.exp(opt_param) # Convert back to linear from log scale
        hparams_vals = jnp.clip(hparams_vals, a_max = server_state.hyper_state.hyperparams.hparam_ub)

        Ep, eta_c, bs = hparams_vals[0], hparams_vals[1], hparams_vals[2]
        hyperparams = HyperParams(
            Ep = Ep,
            eta_c = eta_c,
            bs = bs,
            eta_h = server_state.hyper_state.hyperparams.eta_h,
            alpha = server_state.hyper_state.hyperparams.alpha, 
            hparam_ub = server_state.hyper_state.hyperparams.hparam_ub, 
        )
        hyper_state = HyperState(
            hyperparams = hyperparams,
            init_hparams = server_state.hyper_state.init_hparams,
            opt_state = opt_state,
            opt_param = opt_param,
            hypergrad_glob = hypergrad_glob,
            hypergrad_local = hypergrad_local,
        )        
        return hyper_state

    return federated_algorithm.FederatedAlgorithm(init, apply)

Functions

def create_train_for_each_client(grad_fn: Any, client_optimizer: fedjax.core.optimizers.Optimizer)

Builds client_init, client_step, client_final for for_each_client.

Expand source code
def create_train_for_each_client(grad_fn: Params, client_optimizer: optimizers.Optimizer):
    """Builds client_init, client_step, client_final for for_each_client."""
    normalize_hypergrad = 1

    def client_init(shared_input, client_rng):
        opt_state = client_optimizer.init(shared_input['params'])
        opt_state.hyperparams['learning_rate'] = shared_input['eta_c'] # Already sigmoided
        client_step_state = {
            'params': shared_input['params'],
            'params0': shared_input['params'],
            'opt_state': opt_state,
            'rng': client_rng,
            'step_idx': 0,
            'min_hypergrad': tree_util.tree_l2_norm(shared_input['params']), # Large enough value
            }
        return client_step_state

    def client_step(client_step_state, batch):
        rng, use_rng, metrics_rng = jax.random.split(client_step_state['rng'], num=3)
        grad_opt = grad_fn(client_step_state['params'], batch, use_rng)
        opt_state, params = client_optimizer.apply(grad_opt, client_step_state['opt_state'], client_step_state['params'])
        delta_params = jax.tree_util.tree_multimap(jnp.subtract, client_step_state['params0'], params)
        grad_opt_norm = tree_util.tree_l2_norm(grad_opt)
        delta_params_norm = tree_util.tree_l2_norm(delta_params)
        hypergrad = jnp.where(grad_opt_norm == 0., jnp.array(0),
            jnp.where(delta_params_norm == 0, jnp.array(0),
                jnp.where(normalize_hypergrad == 0 ,
                    fathom.core.tree_util.tree_dot(grad_opt, delta_params),
                    fathom.core.tree_util.tree_dot(grad_opt, delta_params) / grad_opt_norm / delta_params_norm)
            )
        )
        min_hypergrad = jnp.where(hypergrad > client_step_state['min_hypergrad'],
            client_step_state['min_hypergrad'], hypergrad
        )
        next_client_step_state = {
            'params': params,
            'params0': client_step_state['params0'],
            'opt_state': opt_state,
            'rng': rng,
            'step_idx': client_step_state['step_idx'] + 1,
            'min_hypergrad': min_hypergrad,
        }
        return next_client_step_state

    def client_final(shared_input, client_step_state) -> Tuple[Params, jnp.ndarray]:
        delta_params = jax.tree_util.tree_multimap(jnp.subtract, shared_input['params'], client_step_state['params'])
        return delta_params, client_step_state['min_hypergrad'], client_step_state['step_idx']

    return for_each_client.for_each_client(client_init, client_step, client_final)
def estimate_grad_glob(server_state: ServerState, mean_delta_params: Any) ‑> Any
Expand source code
@jax.jit
def estimate_grad_glob(server_state: ServerState, mean_delta_params: Params) -> Params:
    grad_glob = tree_util.tree_weight(server_state.grad_glob, server_state.hyper_state.hyperparams.alpha)
    delta_params = tree_util.tree_weight(mean_delta_params, 1. - server_state.hyper_state.hyperparams.alpha)
    grad_glob = tree_util.tree_add(grad_glob, delta_params)
    # grad_glob = fathom.core.tree_util.tree_inverse_weight(
    #     grad_glob, 
    #     (1. - server_state.hyper_state.hyperparams.alpha ** server_state.round_index)
    # )
    return grad_glob
def federated_averaging(grad_fn: Callable[[Any, Mapping[str, jax._src.numpy.lax_numpy.ndarray], jax._src.numpy.lax_numpy.ndarray], Any], client_optimizer: fedjax.core.optimizers.Optimizer, server_optimizer: fedjax.core.optimizers.Optimizer, hyper_optimizer: fedjax.core.optimizers.Optimizer, client_batch_hparams: fedjax.core.client_datasets.ShuffleRepeatBatchHParams, fathom_init_hparams: HyperParams) ‑> fedjax.core.federated_algorithm.FederatedAlgorithm

Builds federated averaging.

Args

grad_fn
A function from (params, batch_example, rng) to gradients. This can be created with :func:fedjax.core.model.model_grad.
client_optimizer
Optimizer for local client training.
server_optimizer
Optimizer for server update.
client_batch_hparams
Hyperparameters for batching client dataset for train.

Returns

FederatedAlgorithm

Expand source code
def federated_averaging(
        grad_fn: Callable[[Params, BatchExample, PRNGKey], Grads],
        client_optimizer: optimizers.Optimizer,
        server_optimizer: optimizers.Optimizer,
        hyper_optimizer: optimizers.Optimizer,
        client_batch_hparams: client_datasets.ShuffleRepeatBatchHParams,
        fathom_init_hparams: HyperParams,
) -> federated_algorithm.FederatedAlgorithm:
    """Builds federated averaging.

    Args:
        grad_fn: A function from (params, batch_example, rng) to gradients.
            This can be created with :func:`fedjax.core.model.model_grad`.
        client_optimizer: Optimizer for local client training.
        server_optimizer: Optimizer for server update.
        client_batch_hparams: Hyperparameters for batching client dataset for train.

    Returns:
        FederatedAlgorithm
    """
    train_for_each_client = create_train_for_each_client(grad_fn, client_optimizer)

    def server_reset(params: Params, init_hparams: HyperParams) -> ServerState:
        opt_state_server = server_optimizer.init(params)
        opt_param_hyper = jnp.log(jnp.array([init_hparams.Ep, init_hparams.eta_c, init_hparams.bs]))
        opt_state_hyper = hyper_optimizer.init(opt_param_hyper)
        hyper_state = HyperState(
            hyperparams = init_hparams,
            init_hparams = init_hparams,
            opt_state = opt_state_hyper,
            opt_param = opt_param_hyper,
            hypergrad_glob = 0.,
            hypergrad_local = 0.,
        )
        # Need to initialize round_index to 1 for bias comp
        return ServerState(
            params = params, 
            params_bak = params,
            opt_state = opt_state_server, 
            round_index = 1, 
            grad_glob = tree_util.tree_zeros_like(params),
            hyper_state = hyper_state,
        )        

    def init(params: Params) -> ServerState:
        return server_reset(params, fathom_init_hparams)

    def apply(
        server_state: ServerState,
        clients: Sequence[Tuple[
            federated_data.ClientId, 
            client_datasets.ClientDataset, 
            PRNGKey
        ]],
    ) -> Tuple[ServerState, Mapping[federated_data.ClientId, Any]]:
        client_num_examples = {cid: len(cds) for cid, cds, _ in clients}
        Ep: float = server_state.hyper_state.hyperparams.Ep
        bs: int = max(int(server_state.hyper_state.hyperparams.bs + 0.5), 1) if server_state.hyper_state.hyperparams.bs > 0 else -1
        eta_c: float = server_state.hyper_state.hyperparams.eta_c
        batch_clients = [(cid, cds.shuffle_repeat_batch(
            client_datasets.ShuffleRepeatBatchHParams(
                batch_size = bs if bs > 0 else len(cds), 
                num_steps = max(int(jnp.ceil(Ep * client_num_examples[cid] / bs)), 1),
                num_epochs = None, # This is required.  See ShuffleRepeatBatchView implementation in fedjax.core.client_datasets.py.
                drop_remainder = client_batch_hparams.drop_remainder,
                seed = client_batch_hparams.seed,
                skip_shuffle = client_batch_hparams.skip_shuffle,
            )
        ), crng) for cid, cds, crng in clients]
        shared_input = {'params': server_state.params, 'eta_c': eta_c, 'Ep': Ep, 'bs': bs}
        client_diagnostics = {}
        # Running weighted mean of client updates. We do this iteratively to avoid
        # loading all the client outputs into memory since they can be prohibitively
        # large depending on the model parameters size.
        delta_params_sum = tree_util.tree_zeros_like(server_state.params)
        hypergrad_local_sum = jnp.array(0.0)
        num_examples_sum = 0.
        for client_id, (delta_params, min_hypergrad, num_steps) in train_for_each_client(shared_input, batch_clients):
            # Server collecting stats before sending them to server_update for updating params and metrics.
            num_examples = client_num_examples[client_id]
            weighted_delta_params =  tree_util.tree_weight(delta_params, num_examples)
            delta_params_sum = tree_util.tree_add(delta_params_sum, weighted_delta_params)
            hypergrad_local_sum = hypergrad_local_sum + min_hypergrad * num_examples
            num_examples_sum += num_examples
            # We record the l2 norm of client updates as an example, but it is not
            # required for the algorithm.
            client_diagnostics[client_id] = {
                'delta_l2_norm': tree_util.tree_l2_norm(delta_params),
            }
        mean_delta_params = fathom.core.tree_util.tree_inverse_weight(delta_params_sum, num_examples_sum)
        mean_hypergrad_local = hypergrad_local_sum / num_examples_sum
        server_state = server_update(
            server_state = server_state, 
            mean_delta_params = mean_delta_params, 
            hypergrad_local = mean_hypergrad_local, 
        )
        return server_state, client_diagnostics

    def server_update(
        server_state: ServerState, 
        mean_delta_params: Params, 
        hypergrad_local: jnp.ndarray, 
    ) -> ServerState:
        opt_state_server, params = server_optimizer.apply(
            mean_delta_params, 
            server_state.opt_state, 
            server_state.params,
        )
        grad_glob: Params = estimate_grad_glob(server_state, mean_delta_params)
        hyper_state: HyperState = hyper_update(
            server_state = server_state, 
            params = params, 
            delta_params = mean_delta_params,
            hypergrad_local = hypergrad_local,
        )
        return ServerState(
            params = params,
            params_bak = server_state.params_bak, # Keep initial params rather than update
            opt_state = opt_state_server,
            round_index = server_state.round_index + 1,
            hyper_state = hyper_state,
            grad_glob = grad_glob,
        )

    def hyper_update(
        server_state: ServerState,
        params: Params,
        delta_params: Params,
        hypergrad_local: jnp.ndarray,
    ) -> HyperState:

        opt_param, opt_state = server_state.hyper_state.opt_param, server_state.hyper_state.opt_state
        # Do not use the most current grad_glob as the result will bias positive
        hypergrad_glob: float = fathom.core.tree_util.tree_dot(server_state.grad_glob, delta_params)
        # Normalizing hypergrad_global here
        # hypergrad_local already normalized from local calculations
        grad_glob_norm = tree_util.tree_l2_norm(server_state.grad_glob)
        delta_params_norm = tree_util.tree_l2_norm(delta_params)
        hypergrad_glob = jnp.where((grad_glob_norm > 0. and delta_params_norm > 0.),
            hypergrad_glob / grad_glob_norm / delta_params_norm,
            0
        )
        hypergrad = - jnp.array([
            hypergrad_glob + hypergrad_local,   # Ep
            hypergrad_glob,                     # eta_c
            -hypergrad_local,                   # bs
        ]) 
        # This is where individual learning rates are applied, assuming opt is SGD.
        # With any other opt, individual learning rates need to be set at opt instantiation.
        hypergrad = hypergrad * server_state.hyper_state.hyperparams.eta_h

        # EGN gradients are already normalized
        opt_state, opt_param = hyper_optimizer.apply(hypergrad, opt_state, opt_param)
        hparams_vals = jnp.exp(opt_param) # Convert back to linear from log scale
        hparams_vals = jnp.clip(hparams_vals, a_max = server_state.hyper_state.hyperparams.hparam_ub)

        Ep, eta_c, bs = hparams_vals[0], hparams_vals[1], hparams_vals[2]
        hyperparams = HyperParams(
            Ep = Ep,
            eta_c = eta_c,
            bs = bs,
            eta_h = server_state.hyper_state.hyperparams.eta_h,
            alpha = server_state.hyper_state.hyperparams.alpha, 
            hparam_ub = server_state.hyper_state.hyperparams.hparam_ub, 
        )
        hyper_state = HyperState(
            hyperparams = hyperparams,
            init_hparams = server_state.hyper_state.init_hparams,
            opt_state = opt_state,
            opt_param = opt_param,
            hypergrad_glob = hypergrad_glob,
            hypergrad_local = hypergrad_local,
        )        
        return hyper_state

    return federated_algorithm.FederatedAlgorithm(init, apply)

Classes

class HyperParams (eta_c: float, Ep: float, bs: float, alpha: float, eta_h: jax._src.numpy.lax_numpy.ndarray, hparam_ub: jax._src.numpy.lax_numpy.ndarray)

HyperParams(eta_c: float, Ep: float, bs: float, alpha: float, eta_h: jax._src.numpy.lax_numpy.ndarray, hparam_ub: jax._src.numpy.lax_numpy.ndarray)

Expand source code
class HyperParams:
    eta_c: float                # Local learning rate or step size
    Ep: float                  # Number of local steps = ceil(local_samples / batch_size) * Ep, where Ep ~ num epochs worth of data.
    bs: float                   # Local batch size
    alpha: float                # Momentum for glob grad estimation
    eta_h: jnp.ndarray          # Hyper optimizer learning rates for Ep, eta_c, and bs, respectively
    hparam_ub: jnp.ndarray      # Upperbound vals for Ep, eta_c, and bs, respectively.

Class variables

var Ep : float
var alpha : float
var bs : float
var eta_c : float
var eta_h : jax._src.numpy.lax_numpy.ndarray
var hparam_ub : jax._src.numpy.lax_numpy.ndarray

Methods

def replace(self, **updates)

"Returns a new object replacing the specified fields with new values.

Expand source code
def replace(self, **updates):
  """"Returns a new object replacing the specified fields with new values."""
  return dataclasses.replace(self, **updates)
class HyperState (hyperparams: HyperParams, init_hparams: HyperParams, opt_state: Union[jax._src.numpy.lax_numpy.ndarray, Iterable[ForwardRef('ArrayTree')], Mapping[Any, ForwardRef('ArrayTree')]], opt_param: jax._src.numpy.lax_numpy.ndarray, hypergrad_glob: float, hypergrad_local: float)

HyperState(hyperparams: fathom.algorithms.fathom_fedavg.HyperParams, init_hparams: fathom.algorithms.fathom_fedavg.HyperParams, opt_state: Union[jax._src.numpy.lax_numpy.ndarray, Iterable[ForwardRef('ArrayTree')], Mapping[Any, ForwardRef('ArrayTree')]], opt_param: jax._src.numpy.lax_numpy.ndarray, hypergrad_glob: float, hypergrad_local: float)

Expand source code
class HyperState:
    hyperparams: HyperParams 
    init_hparams: HyperParams
    opt_state: optimizers.OptState
    opt_param: jnp.ndarray
    hypergrad_glob: float
    hypergrad_local: float

Class variables

var hypergrad_glob : float
var hypergrad_local : float
var hyperparamsHyperParams
var init_hparamsHyperParams
var opt_param : jax._src.numpy.lax_numpy.ndarray
var opt_state : Union[jax._src.numpy.lax_numpy.ndarray, Iterable[ArrayTree], Mapping[Any, ArrayTree]]

Methods

def replace(self, **updates)

"Returns a new object replacing the specified fields with new values.

Expand source code
def replace(self, **updates):
  """"Returns a new object replacing the specified fields with new values."""
  return dataclasses.replace(self, **updates)
class ServerState (params: Any, params_bak: Any, opt_state: Union[jax._src.numpy.lax_numpy.ndarray, Iterable[ForwardRef('ArrayTree')], Mapping[Any, ForwardRef('ArrayTree')]], round_index: int, hyper_state: HyperState, grad_glob: Any)

State of server passed between rounds.

Attributes

params
A pytree representing the server model parameters.
opt_state
A pytree representing the server optimizer state.
Expand source code
class ServerState:
    """State of server passed between rounds.

    Attributes:
        params: A pytree representing the server model parameters.
        opt_state: A pytree representing the server optimizer state.
    """
    params: Params
    params_bak: Params
    opt_state: optimizers.OptState
    round_index: int
    hyper_state: HyperState
    grad_glob: Params

Class variables

var grad_glob : Any
var hyper_stateHyperState
var opt_state : Union[jax._src.numpy.lax_numpy.ndarray, Iterable[ArrayTree], Mapping[Any, ArrayTree]]
var params : Any
var params_bak : Any
var round_index : int

Methods

def replace(self, **updates)

"Returns a new object replacing the specified fields with new values.

Expand source code
def replace(self, **updates):
  """"Returns a new object replacing the specified fields with new values."""
  return dataclasses.replace(self, **updates)