chop.passes.transform.pruning

chop.passes.transform.pruning#

prune_transform_pass#

chop.passes.graph.transforms.pruning.prune.prune_transform_pass(graph, pass_args: dict = {})[source]#

Apply pruning transformation to the given graph. This is achieved by adding a register_parametrization hook to weights and a register_pre_forward hook to activations

Parameters:

  • graph (MaseGraph) – The input graph to be pruned.

  • pass_args (dict) – Optional arguments for the pruning transformation.

pass_args should take the following form:

pass_config = {
    "weight" : {
        "scope": "local", # ["local, "global"] are available
        "granularity": "element", # ["element"] are available
        "method": "l1", # ["l1", "random"] are available
        "sparsity": 0.5, # a float between 0.0 and 1.0
    },
    "activation" : {
        "scope": "local", # ["local, "global"] are available
        "granularity": "element", # ["element"] are available
        "method": "l1", # ["l1", "random"] are available
        "sparsity": 0.5, # a float between 0.0 and 1.0
}
Returns:

The pruned graph and an empty dictionary.

Return type:

tuple

prune_detach_hook_transform_pass#

chop.passes.graph.transforms.pruning.prune_detach_hook.prune_detach_hook_transform_pass(graph, pass_args: dict = {})[source]#

Apply pruning transformation to the given graph. This is achieved by adding a register_parametrization hook to weights and a register_pre_forward hook to activations

Parameters:

graph (MaseGraph) – The input graph to be pruned.

pass_args can be None or an empty dictionary.

Parameters:

pass_args (dict) – Optional arguments for the pruning transformation.

Returns:

The pruned graph and an empty dictionary.

Return type:

tuple