neural_compressor.pruner.pruner_legacy.gradient_sensitivity

Gradient sensitivity pruner.

Module Contents

Classes

GradientSensitivityPruner

Gradient sensitivity pruner class.

class neural_compressor.pruner.pruner_legacy.gradient_sensitivity.GradientSensitivityPruner(model, local_config, global_config)

Bases: neural_compressor.pruner.pruner_legacy.pruner.Pruner

Gradient sensitivity pruner class.

Parameters:

• model (object) – The original model (currently PyTorchModel instance).

• local_config (Conf) – configs specific for this pruning instance.

• global_config (Conf) – global configs which may be overwritten by local_config.

on_epoch_begin(epoch)

Be called on the beginning of epochs.

on_step_begin(batch_id)

Multiple mask to weight elementwisely when elementwise_prune is True.

on_epoch_end()

Be called on the end of epochs.

parse_weight_name(weight_name_pattern)

Parse weight name.

on_step_end()

Update importance tensor.

compute_mask()

Compute masks according to absolute values.

prune_weight(model, importance, weight_name, parameters)

Prune the specified weight by importance.

update_importance_elementwise(model, importance, weight_name)

Update importance tensor elementwisely.

update_importance_abs(model, importance, weight_name, parameters)

Update importance tensor with absolute gradient.

update_importance_weighted(model, importance, weight_name, parameters)

Update importance tensor with weighted gradient.

prune_by_importance(tensor, importance, num_instances, stride)

Reserve specified number of weights only by importance.