#!/usr/bin/env python3
# *_* coding: utf-8 *_*
# Copyright 2022 Ant Group Co., Ltd.
#
# 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
#
# https://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.
import copy
from typing import Callable, Tuple
import numpy as np
from secretflow.device import PYUObject, proxy
from secretflow.ml.nn.fl.backend.torch.fl_base import BaseTorchModel
from secretflow.ml.nn.fl.backend.torch.utils import TorchModel
from secretflow.ml.nn.fl.sparse import SCRSparse
from secretflow.ml.nn.fl.strategy_dispatcher import register_strategy
[文档]class FedSCR(BaseTorchModel):
"""
FedSCR: A structure-wise aggregation method to identify and remove redundant updates,
it aggregates parameter updates over a particular structure (e.g., filters and channels).
If the sum of the absolute updates of a model structure is lower than a given threshold,
FedSCR will treat the updates in this structure as less important and filter them out.
"""
[文档] def __init__(self, builder_base: Callable[[], TorchModel]):
super().__init__(builder_base)
self._res = []
[文档] def train_step(
self,
updates: np.ndarray,
cur_steps: int,
train_steps: int,
**kwargs,
) -> Tuple[np.ndarray, int]:
"""Accept ps model params,then do local train
Args:
updates: global updates from params server
cur_steps: current train step
train_steps: local training steps
kwargs: strategy-specific parameters
threshold: user-defined threshold, controlling the selectivity of weight updates, filtering insignificant updates
Returns:
Parameters after local training
"""
assert self.model is not None, "Model cannot be none, please give model define"
dp_strategy = kwargs.get('dp_strategy', None)
# prepare for the SCR compression
threshold = kwargs.get('threshold', 0.0)
compressor = SCRSparse(threshold=threshold)
# update current weights
if updates is not None:
weights = [np.add(w, u) for w, u in zip(self.model_weights, updates)]
self.model.update_weights(weights)
num_sample = 0
logs = {}
# store current weights for residual computing
self.model_weights = self.model.get_weights(return_numpy=True)
for _ in range(train_steps):
self.optimizer.zero_grad()
iter_data = next(self.train_iter)
if len(iter_data) == 2:
x, y = iter_data
s_w = None
elif len(iter_data) == 3:
x, y, s_w = iter_data
num_sample += x.shape[0]
y_t = y.argmax(dim=-1)
y_pred = self.model(x)
# do back propagation
loss = self.loss(y_pred, y)
loss.backward()
self.optimizer.step()
for m in self.metrics:
m.update(y_pred, y_t)
loss = loss.item()
logs['train-loss'] = loss
self.logs = self.transform_metrics(logs)
self.epoch_logs = copy.deepcopy(self.logs)
# do SCR compression
if self._res:
client_updates = [
np.add(np.subtract(new_w, old_w), res_u)
for new_w, old_w, res_u in zip(
self.model.get_weights(return_numpy=True),
self.model_weights,
self._res,
)
]
else:
# initial training res is zero
client_updates = [
np.subtract(new_w, old_w)
for new_w, old_w in zip(
self.model.get_weights(return_numpy=True), self.model_weights
)
]
# DP operation
if dp_strategy is not None:
if dp_strategy.model_gdp is not None:
client_updates_tensor = dp_strategy.model_gdp(client_updates)
client_updates = [
client_updates_tensor[i] for i in range(len(client_updates))
]
# do sparsity, filter out minor updates
sparse_client_updates = compressor(client_updates)
# compute new residual
self._res = [
np.subtract(dense_u, sparse_u)
for dense_u, sparse_u in zip(client_updates, sparse_client_updates)
]
self.model.update_weights(self.model_weights)
return sparse_client_updates, num_sample
[文档]@register_strategy(strategy_name='fed_scr', backend='torch')
@proxy(PYUObject)
class PYUFedSCR(FedSCR):
pass