igc.naive.IntGradMeanStd#

class igc.naive.IntGradMeanStd(module, dataset=None, dtld_kwargs=None, forward_method_name=None, forward_method_kwargs=None, dtype=torch.float32, dtype_cat=torch.int32)[source]#

Bases: IntegratedGradients

Mean/std of the IG distribution over the dataset (IGms).

add_embedding_method(embedding_method_name, embedding_method_kwargs=None, embedding_n_cat=None)[source]#

Add an embedding method to preprocess categorical inputs.

Warning

This effect of this method must not excluded from the forward method defined by forward_method_name at initialization.

Parameters:

  • embedding_method_name (str) – Name of the embedding method of the module.

  • embedding_method_kwargs (dict) – Additional keyword arguments to the embedding method of the module.

  • embedding_n_cat (int) – Number of categorical inputs. If None, this value is inferred from the input data types (torch.int16, torch.int32, torch.int64).

Return type:

self

add_ig_post_function(ig_post_func, ig_post_func_kwargs=None)[source]#

Add a function to postprocess individual IG attributions.

Warning

The IG postprocessing function must have the following signature:

def ig_post(ig_x_1, ..., ig_x_n, x_1, ..., x_n, **kwargs)
    # Do something on IG data
    return ig_x_1, ..., ig_x_n
Parameters:

  • ig_post_func (function) – Function to postprocess individual IG attributions.

  • ig_post_func_kwargs (dict) – Additional keyword arguments to the IG postprocessing function.

Return type:

self

compute(x_0=None, y_idx=None, n_steps=64, batch_size=None, x_seed=None, x_0_seed=100, n_x=None, check_error=True)[source]#

Compute mean/std of the IG distribution over the dataset (IGms).

Parameters:

  • x_0 (None | int | float | ArrayLike | tuple(ArrayLike))

  • y_idx (None | int | ArrayLike)

  • n_steps (int)

  • batch_size (int | tuple(int))

  • x_seed (None | int)

  • x_0_seed (None | int)

  • n_x (None | int)

  • check_error (bool)

Return type:

tuple(ArrayLike | tuple(ArrayLike))