fl4health.utils.random module

generate_hash(length=8)[source]

Generates unique hash used as id for client. NOTE: This generation is unaffected by setting of random seeds.

Parameters:

length (int) – The length of the hash generated. Maximum length is 32

Returns:

hash

Return type:

str

restore_random_state(random_state, numpy_state, torch_state)[source]

Restore the state of the random number generators for Python, NumPy, and PyTorch. This will allow you to restore the state of the random number generators to a previously saved state.

Parameters:

  • random_state (tuple[Any, ...]) – The state of the Python random number generator

  • numpy_state (dict[str, Any]) – The state of the NumPy random number generator

  • torch_state (torch.Tensor) – The state of the PyTorch random number generator

Return type:

None

save_random_state()[source]

Save the state of the random number generators for Python, NumPy, and PyTorch. This will allow you to restore the state of the random number generators at a later time.

Returns:

A tuple containing the state of the random number

generators for Python, NumPy, and

Return type:

tuple[tuple[Any, …], dict[str, Any], torch.Tensor]

set_all_random_seeds(seed=42, use_deterministic_torch_algos=False, disable_torch_benchmarking=False)[source]

Set seeds for python random, numpy random, and pytorch random. It also offers the option to force pytorch to use deterministic algorithm for certain methods and layers see: https://pytorch.org/docs/stable/generated/torch.use_deterministic_algorithms.html) for more details. Finally, it allows one to disable cuda benchmarking, which can also affect the determinism of pytorch training outside of random seeding. For more information on reproducibility in pytorch see: https://pytorch.org/docs/stable/notes/randomness.html

NOTE: If the use_deterministic_torch_algos flag is set to True, you may need to set the environment variable CUBLAS_WORKSPACE_CONFIG to something like :4096:8, to avoid CUDA errors. Additional documentation may be found here: https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility

Parameters:

  • seed (int | None, optional) – The seed value to be used for random number generators. Default is 42. Seed setting will no-op if the seed is explicitly set to None

  • use_deterministic_torch_algos (bool, optional) – Whether or not to set torch.use_deterministic_algorithms to True. Defaults to False.

  • disable_torch_benchmarking (bool, optional) – Whether to explicitly disable cuda benchmarking in torch processes. Defaults to False.

Return type:

None

unset_all_random_seeds()[source]

Set random seeds for Python random, NumPy, and PyTorch to None. Running this function would undo the effects of set_all_random_seeds.

Return type:

None