neural_compressor.compression.pruner.dsnot

Module Contents

Functions

return_reorder_indice(input_tensor)

For instance:

DSnoT(initial_metric, sparsity_ratio, wrapped_layer[, ...])

param pow_of_var_regrowing:

The power of variance, float type, default=1.
neural_compressor.compression.pruner.dsnot.return_reorder_indice(input_tensor)[source]

For instance: [[1., -2., 3.], [-2, 2., -4], [5., 6., -7], [-6, -7, -4]] return indices of [[-2., 3., 1.], [-2., -4., 2.], [-7., 6., 5.], [-6., -7., -4.]] Description: The relative order in the positive number remains unchanged, and the relative order in the negative number is flipped.

neural_compressor.compression.pruner.dsnot.DSnoT(initial_metric, sparsity_ratio, wrapped_layer, prune_n=0, prune_m=0, pow_of_var_regrowing=1.0, max_cycle_time=100, without_same_sign=True, update_threshold=0.1)[source]
Parameters:

  • pow_of_var_regrowing – The power of variance, float type, default=1.

  • max_cycle_time – Max cycle time.

  • without_same_sign –

  • sign (without same) –

  • bool –

  • True. (default) –

  • update_threshold – update threshold, float, default 0.1.

See the original paper: https://arxiv.org/pdf/2310.08915.pdf