p2pfl.learning.compression.topk_strategy module

Top K Sparsification optimization strategy.

class p2pfl.learning.compression.topk_strategy.TopKSparsification[source]

Bases: TensorCompressor

Top-K sparsification.

Keeps only the top k largest values in model parameters.

apply_strategy(params, k=0.1)[source]

Reduces params by taking only the top k ones.

Parameters:

  • params (list[ndarray]) – The parameters to compress.

  • k (float) – Percentage of parameters to keep between 0 and 1.

Return type:

tuple[list[ndarray], dict]

reverse_strategy(params, additional_info)[source]

Decompress params.

Parameters:

  • params (list[ndarray]) – The parameters to decompress.

  • additional_info (dict) – Additional information to decompress.

Return type:

list[ndarray]