A perturbation based approach to compute attribution, involving replacing each contiguous rectangular region with a given baseline / reference, and computing the difference in output. For features located in multiple regions (hyperrectangles), the corresponding output differences are averaged to compute the attribution for that feature.
The first patch is applied with the corner aligned with all indices 0, and strides are applied until the entire dimension range is covered. Note that this may cause the final patch applied in a direction to be cut-off and thus smaller than the target occlusion shape.
More details regarding the occlusion (or grey-box / sliding window) method can be found in the original paper and in the DeepExplain implementation. https://arxiv.org/abs/1311.2901 https://github.com/marcoancona/DeepExplain/blob/master/deepexplain/tensorflow/methods.py#L401
forward_func (callable) – The forward function of the model or any modification of it
attribute(inputs, sliding_window_shapes, strides=None, baselines=None, target=None, additional_forward_args=None, perturbations_per_eval=1)¶
inputs (tensor or tuple of tensors) – Input for which occlusion attributions are computed. If forward_func takes a single tensor as input, a single input tensor should be provided. If forward_func takes multiple tensors as input, a tuple of the input tensors should be provided. It is assumed that for all given input tensors, dimension 0 corresponds to the number of examples (aka batch size), and if multiple input tensors are provided, the examples must be aligned appropriately.
sliding_window_shapes (tuple or tuple of tuples) – Shape of patch (hyperrectangle) to occlude each input. For a single input tensor, this must be a tuple of length equal to the number of dimensions of the input tensor - 1, defining the dimensions of the patch. If the input tensor is 1-d, this should be an empty tuple. For multiple input tensors, this must be a tuple containing one tuple for each input tensor defining the dimensions of the patch for that input tensor, as described for the single tensor case.
strides (int or tuple or tuple of ints or tuple of tuples, optional) – This defines the step by which the occlusion hyperrectangle should be shifted by in each direction for each iteration. For a single tensor input, this can be either a single integer, which is used as the step size in each direction, or a tuple of integers matching the number of dimensions in the occlusion shape, defining the step size in the corresponding dimension. For multiple tensor inputs, this can be either a tuple of integers, one for each input tensor (used for all dimensions of the corresponding tensor), or a tuple of tuples, providing the stride per dimension for each tensor. To ensure that all inputs are covered by at least one sliding window, the stride for any dimension must be <= the corresponding sliding window dimension if the sliding window dimension is less than the input dimension. If None is provided, a stride of 1 is used for each dimension of each input tensor. Default: None
baselines (scalar, tensor, tuple of scalars or tensors, optional) –
Baselines define reference value which replaces each feature when occluded. Baselines can be provided as:
a single tensor, if inputs is a single tensor, with exactly the same dimensions as inputs or broadcastable to match the dimensions of inputs
a single scalar, if inputs is a single tensor, which will be broadcasted for each input value in input tensor.
a tuple of tensors or scalars, the baseline corresponding to each tensor in the inputs’ tuple can be:
either a tensor with matching dimensions to corresponding tensor in the inputs’ tuple or the first dimension is one and the remaining dimensions match with the corresponding input tensor.
or a scalar, corresponding to a tensor in the inputs’ tuple. This scalar value is broadcasted for corresponding input tensor.
In the cases when baselines is not provided, we internally use zero scalar corresponding to each input tensor. Default: None
Output indices for which difference is computed (for classification cases, this is usually the target class). If the network returns a scalar value per example, no target index is necessary. For general 2D outputs, targets can be either:
a single integer or a tensor containing a single integer, which is applied to all input examples
a list of integers or a 1D tensor, with length matching the number of examples in inputs (dim 0). Each integer is applied as the target for the corresponding example.
For outputs with > 2 dimensions, targets can be either:
A single tuple, which contains #output_dims - 1 elements. This target index is applied to all examples.
A list of tuples with length equal to the number of examples in inputs (dim 0), and each tuple containing #output_dims - 1 elements. Each tuple is applied as the target for the corresponding example.
additional_forward_args (any, optional) – If the forward function requires additional arguments other than the inputs for which attributions should not be computed, this argument can be provided. It must be either a single additional argument of a Tensor or arbitrary (non-tuple) type or a tuple containing multiple additional arguments including tensors or any arbitrary python types. These arguments are provided to forward_func in order following the arguments in inputs. For a tensor, the first dimension of the tensor must correspond to the number of examples. For all other types, the given argument is used for all forward evaluations. Note that attributions are not computed with respect to these arguments. Default: None
perturbations_per_eval (int, optional) – Allows multiple occlusions to be included in one batch (one call to forward_fn). By default, perturbations_per_eval is 1, so each occlusion is processed individually. Each forward pass will contain a maximum of perturbations_per_eval * #examples samples. For DataParallel models, each batch is split among the available devices, so evaluations on each available device contain at most (perturbations_per_eval * #examples) / num_devices samples. Default: 1
- attributions (tensor or tuple of tensors):
The attributions with respect to each input feature. Attributions will always be the same size as the provided inputs, with each value providing the attribution of the corresponding input index. If a single tensor is provided as inputs, a single tensor is returned. If a tuple is provided for inputs, a tuple of corresponding sized tensors is returned.
- Return type
tensor or tuple of tensors of attributions
>>> # SimpleClassifier takes a single input tensor of size Nx4x4, >>> # and returns an Nx3 tensor of class probabilities. >>> net = SimpleClassifier() >>> # Generating random input with size 2 x 4 x 4 >>> input = torch.randn(2, 4, 4) >>> # Defining Occlusion interpreter >>> ablator = Occlusion(net) >>> # Computes occlusion attribution, ablating each 3x3 patch, >>> # shifting in each direction by the default of 1. >>> attr = ablator.attribute(input, target=1, sliding_window_shapes=(3,3))