arguments.txt

path: The path to evaluate a pretrained model. 

wd_factor: The weight decay factor for training

normalization: Type of initial normalization for model parameters.

epochs: number of the experiment total epochs

clients: number of the experiment clients

t_round: number of the local epoch of each client or each round

lr: the learning rate of the experiment

dataset: the dataset to use (right now it supports only cifar10 and cifar100)

mod: which model to use for training (right now only vgg and resnet models are supported)

depth: the depth of the model (i.e. 56 for resnet56)

layer: the layer to use during distillation (last and prelast layer are supported)

mode: "distillation" for federated distillation and traditional for the federated learning scheme

common dataset size: the size of the dataset that it is common to all clients (as percentage of the total training dataset)

coef_t, coef_d: weights of the losses during training or distillation

external distillation: the path to a pretrained teacher model employed for distillation

average: whether or not to average local anchors before transmit them to central node

loss: which loss to use for distillation (mse, crossentropy or kl loss)

temperature: denotes the weighting of the contrastive loss terms

exemplar_percentage: the size of the dataset that it is employed as exemplar set (as percentage of the total training dataset)

incremental_step: number of new classes each new task introduces

incremental_rounds: number of federated rounds of each incremental task

model_buffer_size: the number of models of previous incremental rounds employed to the contrastive loss

mu: weight factor for the contrastive loss

beta: the concentration parameter of the Dirichlet distriution. It controls the non-IIDnes of data (applicable only if partition=noniid)

multiloss: whether or not to use intermediate multi-scale knowledge distillation losses

multiloss type: the multiloss type as described at section 4.3.1 of the paper

scale: the scale of the multi-scale knowledge distillation loss