Qfedavg Strategy

qFedAvg¶

According to the article FAIR RESOURCE ALLOCATION IN FEDERATED LEARNING, we implement qfedavg aggregation algorithm.

The complete source code reference for this example qFedAvg.

To start a qfedav server¶

strategy = message_type.STRATEGY_qFEDAVG # define server type

server = AggregateServer(args.addr, strategy, args.num, q=.2, learning_rate = 1)
server.run()

or

python iflearner/business/homo/aggregate_server.py -n 2 --strategy qFedAvg   --strategy_params {"q":.2, "learning_rate":1}

To start a client¶

See how to use

It should be noted that before the client starts fitting, you need to obtain the loss value loss of the current model on the training data. Then, you should override the Trainer.get method on the client side and add the loss keyword to the uploaded parameters.

def evaluate_traindata(self):
    batch_time = AverageMeter("Time", ":6.3f", Summary.AVERAGE)
    losses = AverageMeter("Loss", ":.4e", Summary.AVERAGE)
    top1 = AverageMeter("Acc@1", ":6.2f", Summary.AVERAGE)
    top5 = AverageMeter("Acc@5", ":6.2f", Summary.AVERAGE)
    progress = ProgressMeter(
        len(self._train_loader), [batch_time, losses, top1, top5], prefix="Test on training data: "
    )

    with torch.no_grad():
        end = time.time()
        for i, (images, target) in enumerate(self._train_loader):
            if self._args.gpu is not None:
                images = images.cuda(self._args.gpu, non_blocking=True)
            if torch.cuda.is_available():
                target = target.cuda(self._args.gpu, non_blocking=True)

            # compute output
            output = self._model(images)
            loss = self._criterion(output, target)

            # measure accuracy and record loss
            acc1, acc5 = accuracy(output, target, topk=(1, 5))
            losses.update(loss.item(), images.size(0))
            top1.update(acc1[0], images.size(0))
            top5.update(acc5[0], images.size(0))

            # measure elapsed time
            batch_time.update(time.time() - end)
            end = time.time()

            if i % self._args.print_freq == 0:
                progress.display(i)

        progress.display_summary()
    self._fs = losses.avg

def get(self, param_type=''):
    parameters = dict()
    parameters['loss'] = np.array([self._fs])
    for name, p in self._model.named_parameters():
        if p.requires_grad:
            parameters[name.replace('module.', '')
                        ] = p.cpu().detach().numpy()

    return parameters