chop.passes.graph.pruning

add_pruning_metadata_analysis_pass

chop.passes.graph.analysis.pruning.calculate_sparsity.add_pruning_metadata_analysis_pass(graph, pass_args: dict = {})

Add post-pruning metadata analysis pass to the given graph, the graph must have been pruned.

Parameters:

• graph (MaseGraph) – The MaseGraph to which the pruning metadata analysis pass will be added.

• pass_args – Additional arguments for the pruning metadata analysis pass.

This pass requires a dummy_in and a bool value for add_value. If add value is true, the mask values would be added to meta data. :type pass_args: dict

pass_args = {
    'dummy_in': torch.tensor(...),
    'add_meta': True # bool value whether to add the natural sparsity to the metadata
}

Returns:

The updated graph and sparsity information.

The returned dict contains {‘weight_sparsity’: float, ‘activation_sparsity’: float} :rtype: tuple(MaseGraph, dict)

add_natural_sparsity_metadata_analysis_pass

chop.passes.graph.analysis.pruning.calculate_natural_sparsity.add_natural_sparsity_metadata_analysis_pass(graph, pass_args: dict = {})

Add natural sparsity metadata analysis pass to the given MaseGraph. This is normally used to inspect on the natural sparsity values on both weights and activations.

Parameters:

• graph (MaseGraph) – The MaseGraph to which the analysis pass will be added.

• pass_args – Additional arguments for the analysis pass.

{‘dummy_in’: tensor, ‘add_meta’ bool}, add_meta controls whether he natural sparsity would be registered in mase metadata. :type pass_args: dict

pass_args = {
    'dummy_in': torch.tensor(...),
    'add_meta': True # bool value whether to add the natural sparsity to the metadata
}

Returns:

The updated MaseGraph and sparsity information.

The returned dict contains {name (str): sparsity_value (float)} :rtype: tuple

Examples:

A sample output dict:

{
    'avg_activation': tensor(0.6709),
    'avg_weight': tensor(1.0000),
    'conv1_activation': (6144, tensor(1.)),
    'conv1_weight': (9408, tensor(1.)),
    'fc_activation': (1024, tensor(0.6289)),
    'fc_weight': (5120, tensor(1.)),
    'layer1.0.conv1_activation': (8192, tensor(0.8810)),
    'layer1.0.conv1_weight': (36864, tensor(1.)),
    ...
}

hook_inspection_analysis_pass

chop.passes.graph.analysis.pruning.hook_inspector.hook_inspection_analysis_pass(graph, pass_args: dict = {})

Remove and provide hook information of the modules.

Parameters:

• graph (MaseGraph) – The MaseGraph to which the pruning metadata analysis pass will be added.

• pass_args (dict) – Additional arguments for the pruning metadata analysis pass. This pass does not need any values, so an empty dictionary is fine

pass_args can be None or an empty dictionary.

Returns:

The updated graph and sparsity information. The returned dict contains {‘module_name’: (hook_id, hook_fn)}

Return type:

tuple(MaseGraph, dict)

Examples:

A sample output dict:

{
    'feature_layers.0_0': (
        0, <function get_activation_hook.<locals>.sparsify_input at 0x7f9544528c10>),
    'feature_layers.3_1': (
        1, <function get_activation_hook.<locals>.sparsify_input at 0x7f9544528ca0>),
    'feature_layers.7_2': (
        2, <function get_activation_hook.<locals>.sparsify_input at 0x7f9544528d30>),
}