A perturbation based approach to computing attribution, involving replacing each input feature with a given baseline / reference, and computing the difference in output. By default, each scalar value within each input tensor is taken as a feature and replaced independently. Passing a feature mask, allows grouping features to be ablated together. This can be used in cases such as images, where an entire segment or region can be ablated, measuring the importance of the segment (feature group). Each input scalar in the group will be given the same attribution value equal to the change in target as a result of ablating the entire feature group.
The forward function can either return a scalar per example or a tensor of a fixed sized tensor (or scalar value) for the full batch, i.e. the output does not grow as the batch size increase. If the output is fixed we consider this model to be an “aggregation” of the inputs. In the fixed sized output mode we require perturbations_per_eval == 1 and the feature_mask to be either None or for all of them to have 1 as their first dimension (i.e. a feature mask requires to be applied to all inputs).
forward_func (callable) – The forward function of the model or any modification of it
attribute(inputs, baselines=None, target=None, additional_forward_args=None, feature_mask=None, perturbations_per_eval=1, **kwargs)¶
inputs (tensor or tuple of tensors) – Input for which ablation 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.
baselines (scalar, tensor, tuple of scalars or tensors, optional) –
Baselines define reference value which replaces each feature when ablated. 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 gradients are 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
feature_mask (tensor or tuple of tensors, optional) – feature_mask defines a mask for the input, grouping features which should be ablated together. feature_mask should contain the same number of tensors as inputs. Each tensor should be the same size as the corresponding input or broadcastable to match the input tensor. Each tensor should contain integers in the range 0 to num_features - 1, and indices corresponding to the same feature should have the same value. Note that features within each input tensor are ablated independently (not across tensors). If the forward function returns a single scalar per batch, we enforce that the first dimension of each mask must be 1, since attributions are returned batch-wise rather than per example, so the attributions must correspond to the same features (indices) in each input example. If None, then a feature mask is constructed which assigns each scalar within a tensor as a separate feature, which is ablated independently. Default: None
perturbations_per_eval (int, optional) – Allows ablation of multiple features to be processed simultaneously in one call to forward_fn. 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. If the forward function’s number of outputs does not change as the batch size grows (e.g. if it outputs a scalar value), you must set perturbations_per_eval to 1 and use a single feature mask to describe the features for all examples in the batch. Default: 1
**kwargs (Any, optional) – Any additional arguments used by child classes of FeatureAblation (such as Occlusion) to construct ablations. These arguments are ignored when using FeatureAblation directly. Default: None
- attributions (tensor or tuple of tensors):
The attributions with respect to each input feature. If the forward function returns a scalar value per example, attributions will be the same size as the provided inputs, with each value providing the attribution of the corresponding input index. If the forward function returns a scalar per batch, then attribution tensor(s) will have first dimension 1 and the remaining dimensions will match the input. If a single tensor is provided as inputs, a single tensor is returned. If a tuple of tensors 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 FeatureAblation interpreter >>> ablator = FeatureAblation(net) >>> # Computes ablation attribution, ablating each of the 16 >>> # scalar input independently. >>> attr = ablator.attribute(input, target=1) >>> # Alternatively, we may want to ablate features in groups, e.g. >>> # grouping each 2x2 square of the inputs and ablating them together. >>> # This can be done by creating a feature mask as follows, which >>> # defines the feature groups, e.g.: >>> # +---+---+---+---+ >>> # | 0 | 0 | 1 | 1 | >>> # +---+---+---+---+ >>> # | 0 | 0 | 1 | 1 | >>> # +---+---+---+---+ >>> # | 2 | 2 | 3 | 3 | >>> # +---+---+---+---+ >>> # | 2 | 2 | 3 | 3 | >>> # +---+---+---+---+ >>> # With this mask, all inputs with the same value are ablated >>> # simultaneously, and the attribution for each input in the same >>> # group (0, 1, 2, and 3) per example are the same. >>> # The attributions can be calculated as follows: >>> # feature mask has dimensions 1 x 4 x 4 >>> feature_mask = torch.tensor([[[0,0,1,1],[0,0,1,1], >>> [2,2,3,3],[2,2,3,3]]]) >>> attr = ablator.attribute(input, target=1, feature_mask=feature_mask)