Usage¶

Quick Start¶

python main.py --dataset Electricity --strategy FedAvg --model Linear \
    --input_len 96 --output_len 720 --iterations 100

Load defaults from a JSON config file (CLI flags override):

python main.py --config_file configs/my_experiment.json --iterations 200

---

Reference¶

General¶

Flag	Type	Default	Description
--config_file	str	None	JSON config file; CLI flags override
--seed	int	941	Random seed
--times	int	1	Number of independent runs
--prev	int	0	Resume offset (skip first N runs)
--num_workers	int	4	Parallel workers
--device	str	cuda	cpu or cuda
--device_id	str	0	CUDA device id(s), comma-separated
--efficiency	str	high	Device residency — low / med / high
--save_local_model	flag	False	Save each client's local model checkpoint
--keep_useless_run	flag	False	Keep runs interrupted by KeyboardInterrupt
--compact	flag	False	Merge per-seed files and remove intermediates after a successful run

Save Path¶

Flag	Type	Default	Description
--project	str	./runs	Root output directory
--name	str	exp	Experiment name (auto-incremented if exists)
--sep	str	""	Separator for auto-increment suffix

Dataset¶

Flag	Type	Default	Description
--dataset	str	ETDatasetHour	Dataset name — see docs/datasets.md
--input_len	int	96	Lookback window length
--offset_len	int	0	Gap between input and output windows
--output_len	int	96	Forecast horizon
--batch_size	int	32	Batch size
--scaler	str	Standard	Normalisation — BaseScaler / MaxAbs / MinMax / Robust / Standard
--train_ratio	float	0.8	Fraction of each client's data used for training
--sample_ratio	float	1.0	Random subsample ratio applied to each client's train set

Federation¶

Flag	Type	Default	Description
--strategy	str	LocalOnly	FL strategy — see docs/strategies.md
--model	str	DLinear	Model architecture — see docs/models.md
--iterations	int	10	Global federation rounds
--patience	int	0	Early-stopping patience; 0 = disabled
--join_ratio	float	1.0	Fraction of clients selected per round
--random_join_ratio	bool	False	Randomly vary join ratio each round
--eval_gap	int	1	Evaluate every N rounds
--skip_eval_train	flag	False	Skip train-set evaluation each round
--exclude_server_model_processes	flag	False	Disable server-side model saving and summarisation
--return_diff	bool	False	Clients send weight delta instead of full model

Client¶

Flag	Type	Default	Description
--optimizer	str	Adam	Optimizer — see docs/optimizers.md
--learning_rate	float	0.0001	Local learning rate
--epochs	int	1	Local update steps per round
--loss	str	MSE	Loss function — see docs/losses.md
--scheduler	str	BaseScheduler	LR scheduler — see docs/schedulers.md

Adversarial Eval¶

Applies only to sFL-based strategies. In benign mode (defaults) the behaviour is identical to a standard tFL run.

Flag	Type	Default	Description
--attack	str	NoAttack	Attack injected into malicious clients' packages each round
--malicious_frac	float	0.0	Fraction of clients designated as Byzantine; 0 = benign mode

Krum-specific

Flag	Type	Default	Description
--num_malicious_clients	int	0	f assumed by Krum when computing neighbor scores; 0 = derive from --malicious_frac
--num_clients_to_keep	int	0	Multi-Krum: average the top-k lowest-score clients; 0 = classical Krum

FedTrimmedAvg-specific

Flag	Type	Default	Description
--beta	float	0.2	Fraction trimmed from each tail per coordinate

# Krum under Sign-Flip attack, 20% Byzantine clients
python main.py --dataset Electricity --strategy Krum --model DLinear \
    --attack SignFlip --malicious_frac 0.2 --iterations 100

New-Client Onboarding¶

Holds out a fraction of clients from training. After federation ends, each held-out client runs a local adaptation step and is evaluated separately. Results are logged to the server log and saved to new_client_results.json in the run directory.

Flag	Type	Default	Description
--exclude_ratio	float	0.0	Fraction of clients held out; sampled randomly with --seed
--adapt_T	int	None	First T training windows for new clients; None = full train set
--adapt_epochs	int	1	Local adaptation epochs for new clients

Strategies override client.adapt() for custom adaptation logic. The default fine-tunes from the global model with gradient descent. Strategies with no global model (e.g. LocalOLS) override to run their closed-form solve on the T windows instead.

# 20% new clients, each adapted on their first 100 windows
python main.py --dataset Electricity --strategy FedAvg --model Linear \
    --exclude_ratio 0.2 --adapt_T 100 --adapt_epochs 5