Source code for tint.attr.feature_ablation

#!/usr/bin/env python3

import math
from typing import Any, Callable, Tuple, Union, cast

import torch
from captum._utils.common import (
    _expand_additional_forward_args,
    _expand_target,
    _format_additional_forward_args,
    _format_inputs,
    _format_output,
    _is_tuple,
    _run_forward,
)
from captum._utils.progress import progress
from captum._utils.typing import (
    BaselineType,
    TargetType,
    TensorOrTupleOfTensorsGeneric,
)
from captum.attr._utils.attribution import PerturbationAttribution
from captum.attr._utils.common import _format_input_baseline
from captum.log import log_usage
from torch import Tensor, dtype


[docs]class FeatureAblation(PerturbationAttribution): r""" Feature ablation. 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). """ def __init__(self, forward_func: Callable) -> None: r""" Args: forward_func (callable): The forward function of the model or any modification of it """ PerturbationAttribution.__init__(self, forward_func) self.use_weights = False
[docs] @log_usage() def attribute( self, inputs: TensorOrTupleOfTensorsGeneric, baselines: BaselineType = None, target: TargetType = None, additional_forward_args: Any = None, feature_mask: Union[None, Tensor, Tuple[Tensor, ...]] = None, perturbations_per_eval: int = 1, attributions_fn: Callable = None, show_progress: bool = False, **kwargs: Any, ) -> TensorOrTupleOfTensorsGeneric: r""" Args: 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 target (int, tuple, tensor or list, optional): 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. Default: None 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 attributions_fn (Callable, optional): Applies a function to the attributions before performing the weighted sum. Default: None show_progress (bool, optional): Displays the progress of computation. It will try to use tqdm if available for advanced features (e.g. time estimation). Otherwise, it will fallback to a simple output of progress. Default: False **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 Returns: *tensor* or tuple of *tensors* of **attributions**: - **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. Examples:: >>> # 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) """ # Keeps track whether original input is a tuple or not before # converting it into a tuple. is_inputs_tuple = _is_tuple(inputs) inputs, baselines = _format_input_baseline(inputs, baselines) additional_forward_args = _format_additional_forward_args( additional_forward_args ) num_examples = inputs[0].shape[0] feature_mask = ( _format_inputs(feature_mask) if feature_mask is not None else None ) assert ( isinstance(perturbations_per_eval, int) and perturbations_per_eval >= 1 ), "Perturbations per evaluation must be an integer and at least 1." with torch.no_grad(): if show_progress: feature_counts = self._get_feature_counts( inputs, feature_mask, **kwargs ) total_forwards = ( sum( math.ceil(count / perturbations_per_eval) for count in feature_counts ) + 1 ) # add 1 for the initial eval attr_progress = progress( desc=f"{self.get_name()} attribution", total=total_forwards ) attr_progress.update(0) # Computes initial evaluation with all features, which is compared # to each ablated result. initial_eval = _run_forward( self.forward_func, inputs, target, additional_forward_args ) if show_progress: attr_progress.update() agg_output_mode = FeatureAblation._find_output_mode( perturbations_per_eval, feature_mask ) # get as a 2D tensor (if it is not a scalar) if isinstance(initial_eval, torch.Tensor): initial_eval = initial_eval.reshape(1, -1) num_outputs = initial_eval.shape[1] else: num_outputs = 1 if not agg_output_mode: assert ( isinstance(initial_eval, torch.Tensor) and num_outputs == num_examples ), ( "expected output of `forward_func` to have " + "`batch_size` elements for perturbations_per_eval > 1 " + "and all feature_mask.shape[0] > 1" ) # Initialize attribution totals and counts attrib_type = cast( dtype, initial_eval.dtype if isinstance(initial_eval, Tensor) else type(initial_eval), ) total_attrib = [ torch.zeros( (num_outputs,) + input.shape[1:], dtype=attrib_type, device=input.device, ) for input in inputs ] # Weights are used in cases where ablations may be overlapping. if self.use_weights: weights = [ torch.zeros( (num_outputs,) + input.shape[1:], device=input.device ).float() for input in inputs ] # Iterate through each feature tensor for ablation for i in range(len(inputs)): # Skip any empty input tensors if torch.numel(inputs[i]) == 0: continue for ( current_inputs, current_add_args, current_target, current_mask, ) in self._ith_input_ablation_generator( i, inputs, additional_forward_args, target, baselines, feature_mask, perturbations_per_eval, **kwargs, ): # modified_eval dimensions: 1D tensor with length # equal to #num_examples * #features in batch modified_eval = _run_forward( self.forward_func, current_inputs, current_target, current_add_args, ) if show_progress: attr_progress.update() # (contains 1 more dimension than inputs). This adds extra # dimensions of 1 to make the tensor broadcastable with the inputs # tensor. if not isinstance(modified_eval, torch.Tensor): eval_diff = initial_eval - modified_eval else: if not agg_output_mode: assert ( modified_eval.numel() == current_inputs[0].shape[0] ), """expected output of forward_func to grow with batch_size. If this is not the case for your model please set perturbations_per_eval = 1""" eval_diff = ( initial_eval - modified_eval.reshape((-1, num_outputs)) ).reshape( (-1, num_outputs) + (len(inputs[i].shape) - 1) * (1,) ) eval_diff = eval_diff.to(total_attrib[i].device) if self.use_weights: weights[i] += current_mask.float().sum(dim=0) if attributions_fn is not None: eval_diff = attributions_fn(eval_diff) total_attrib[i] += ( eval_diff * current_mask.to(attrib_type) ).sum(dim=0) if show_progress: attr_progress.close() # Divide total attributions by counts and return formatted attributions if self.use_weights: attrib = tuple( single_attrib.float() / weight for single_attrib, weight in zip(total_attrib, weights) ) else: attrib = tuple(total_attrib) _result = _format_output(is_inputs_tuple, attrib) return _result
def _ith_input_ablation_generator( self, i, inputs, additional_args, target, baselines, input_mask, perturbations_per_eval, **kwargs, ): """ This method return an generator of ablation perturbations of the i-th input Returns: ablation_iter (generator): yields each perturbation to be evaluated as a tuple (inputs, additional_forward_args, targets, mask). """ extra_args = {} for key, value in kwargs.items(): # For any tuple argument in kwargs, we choose index i of the tuple. if isinstance(value, tuple): extra_args[key] = value[i] else: extra_args[key] = value input_mask = input_mask[i] if input_mask is not None else None ( min_feature, num_features, input_mask, ) = self._get_feature_range_and_mask( inputs[i], input_mask, **extra_args ) num_examples = inputs[0].shape[0] perturbations_per_eval = min(perturbations_per_eval, num_features) baseline = baselines[i] if isinstance(baselines, tuple) else baselines if isinstance(baseline, torch.Tensor): baseline = baseline.reshape((1,) + baseline.shape) if perturbations_per_eval > 1: # Repeat features and additional args for batch size. all_features_repeated = [ torch.cat([inputs[j]] * perturbations_per_eval, dim=0) for j in range(len(inputs)) ] additional_args_repeated = ( _expand_additional_forward_args( additional_args, perturbations_per_eval ) if additional_args is not None else None ) target_repeated = _expand_target(target, perturbations_per_eval) else: all_features_repeated = list(inputs) additional_args_repeated = additional_args target_repeated = target num_features_processed = min_feature while num_features_processed < num_features: current_num_ablated_features = min( perturbations_per_eval, num_features - num_features_processed ) # Store appropriate inputs and additional args based on batch size. if current_num_ablated_features != perturbations_per_eval: current_features = [ feature_repeated[ 0 : current_num_ablated_features * num_examples ] for feature_repeated in all_features_repeated ] current_additional_args = ( _expand_additional_forward_args( additional_args, current_num_ablated_features ) if additional_args is not None else None ) current_target = _expand_target( target, current_num_ablated_features ) else: current_features = all_features_repeated current_additional_args = additional_args_repeated current_target = target_repeated # Store existing tensor before modifying original_tensor = current_features[i] # Construct ablated batch for features in range num_features_processed # to num_features_processed + current_num_ablated_features and return # mask with same size as ablated batch. ablated_features has dimension # (current_num_ablated_features, num_examples, inputs[i].shape[1:]) # Note that in the case of sparse tensors, the second dimension # may not necessarilly be num_examples and will match the first # dimension of this tensor. current_reshaped = current_features[i].reshape( (current_num_ablated_features, -1) + current_features[i].shape[1:] ) ablated_features, current_mask = self._construct_ablated_input( current_reshaped, input_mask, baseline, num_features_processed, num_features_processed + current_num_ablated_features, **extra_args, ) # current_features[i] has dimension # (current_num_ablated_features * num_examples, inputs[i].shape[1:]), # which can be provided to the model as input. current_features[i] = ablated_features.reshape( (-1,) + ablated_features.shape[2:] ) yield tuple( current_features ), current_additional_args, current_target, current_mask # Replace existing tensor at index i. current_features[i] = original_tensor num_features_processed += current_num_ablated_features def _construct_ablated_input( self, expanded_input, input_mask, baseline, start_feature, end_feature, **kwargs, ): r""" Ablates given expanded_input tensor with given feature mask, feature range, and baselines. expanded_input shape is (`num_features`, `num_examples`, ...) with remaining dimensions corresponding to remaining original tensor dimensions and `num_features` = `end_feature` - `start_feature`. input_mask has same number of dimensions as original input tensor (one less than `expanded_input`), and can have first dimension either 1, applying same feature mask to all examples, or `num_examples`. baseline is expected to be broadcastable to match `expanded_input`. This method returns the ablated input tensor, which has the same dimensionality as `expanded_input` as well as the corresponding mask with either the same dimensionality as `expanded_input` or second dimension being 1. This mask contains 1s in locations which have been ablated (and thus counted towards ablations for that feature) and 0s otherwise. """ current_mask = torch.stack( [input_mask == j for j in range(start_feature, end_feature)], dim=0 ).long() ablated_tensor = ( expanded_input * (1 - current_mask).to(expanded_input.dtype) ) + (baseline * current_mask.to(expanded_input.dtype)) return ablated_tensor, current_mask def _get_feature_range_and_mask(self, input, input_mask, **kwargs): if input_mask is None: # Obtain feature mask for selected input tensor, matches size of # 1 input example, (1 x inputs[i].shape[1:]) input_mask = torch.reshape( torch.arange(torch.numel(input[0]), device=input.device), input[0:1].shape, ).long() return ( torch.min(input_mask).item(), torch.max(input_mask).item() + 1, input_mask, ) def _get_feature_counts(self, inputs, feature_mask, **kwargs): """return the numbers of input features""" if not feature_mask: return tuple( inp[0].numel() if inp.numel() else 0 for inp in inputs ) return tuple( (mask.max() - mask.min()).item() + 1 if mask is not None else (inp[0].numel() if inp.numel() else 0) for inp, mask in zip(inputs, feature_mask) ) @staticmethod def _find_output_mode( perturbations_per_eval: int, feature_mask: Union[None, TensorOrTupleOfTensorsGeneric], ) -> bool: """ Returns True if the output mode is "aggregation output mode" Aggregation output mode is defined as: when there is no 1:1 correspondence with the `num_examples` (`batch_size`) and the amount of outputs your model produces, i.e. the model output does not grow in size as the input becomes larger. We assume this is the case if `perturbations_per_eval == 1` and your feature mask is None or is associated to all examples in a batch (fm.shape[0] == 1 for all fm in feature_mask). """ return perturbations_per_eval == 1 and ( feature_mask is None or all( len(sm.shape) == 0 or sm.shape[0] == 1 for sm in feature_mask ) )