def _next_infidelity(current_n_perturb_samples: int) -> Tensor: perturbations, inputs_perturbed = _generate_perturbations( current_n_perturb_samples ) perturbations = _format_tensor_into_tuples(perturbations) inputs_perturbed = _format_tensor_into_tuples(inputs_perturbed) _validate_inputs_and_perturbations( cast(Tuple[Tensor, ...], inputs), cast(Tuple[Tensor, ...], inputs_perturbed), cast(Tuple[Tensor, ...], perturbations), ) targets_expanded = _expand_target( target, current_n_perturb_samples, expansion_type=ExpansionTypes.repeat_interleave, ) additional_forward_args_expanded = _expand_additional_forward_args( additional_forward_args, current_n_perturb_samples, expansion_type=ExpansionTypes.repeat_interleave, ) inputs_perturbed_fwd = _run_forward( forward_func, inputs_perturbed, targets_expanded, additional_forward_args_expanded, ) inputs_fwd = _run_forward(forward_func, inputs, target, additional_forward_args) inputs_fwd = torch.repeat_interleave( inputs_fwd, current_n_perturb_samples, dim=0 ) inputs_minus_perturb = inputs_fwd - inputs_perturbed_fwd attributions_expanded = tuple( torch.repeat_interleave(attribution, current_n_perturb_samples, dim=0) for attribution in attributions ) attributions_times_perturb = tuple( (attribution_expanded * perturbation).view(attribution_expanded.size(0), -1) for attribution_expanded, perturbation in zip( attributions_expanded, perturbations ) ) attribution_times_perturb_sums = sum( [ torch.sum(attribution_times_perturb, dim=1) for attribution_times_perturb in attributions_times_perturb ] ) return torch.sum( torch.pow( attribution_times_perturb_sums - inputs_minus_perturb.view(-1), 2 ).view(bsz, -1), dim=1, )
def _expand_inputs_baselines_targets( self, baselines: Tuple[Tensor, ...], inputs: Tuple[Tensor, ...], target: TargetType, additional_forward_args: Any, ) -> Tuple[Tuple[Tensor, ...], Tuple[Tensor, ...], TargetType, Any]: inp_bsz = inputs[0].shape[0] base_bsz = baselines[0].shape[0] expanded_inputs = tuple([ input.repeat_interleave(base_bsz, dim=0).requires_grad_() for input in inputs ]) expanded_baselines = tuple([ baseline.repeat((inp_bsz, ) + tuple([1] * (len(baseline.shape) - 1))).requires_grad_() for baseline in baselines ]) expanded_target = _expand_target( target, base_bsz, expansion_type=ExpansionTypes.repeat_interleave) input_additional_args = (_expand_additional_forward_args( additional_forward_args, base_bsz, expansion_type=ExpansionTypes.repeat_interleave, ) if additional_forward_args is not None else None) return ( expanded_inputs, expanded_baselines, expanded_target, input_additional_args, )
def attribute( self, inputs: Union[Tensor, Tuple[Tensor, ...]], baselines: BaselineType = None, target: TargetType = None, additional_forward_args: Any = None, return_convergence_delta: bool = False, attribute_to_layer_input: bool = False, custom_attribution_func: Union[None, Callable[..., Tuple[Tensor, ...]]] = None, ) -> Union[Tensor, Tuple[Tensor, ...], Tuple[Union[Tensor, Tuple[ Tensor, ...]], Tensor]]: r""" Args: inputs (tensor or tuple of tensors): Input for which layer 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 samples that are compared with the inputs. In order to assign attribution scores DeepLift computes the differences between the inputs/outputs and corresponding references. Baselines can be provided as: - a single tensor, if inputs is a single tensor, with exactly the same dimensions as inputs or the first dimension is one and the remaining dimensions match with 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. Note that attributions are not computed with respect to these arguments. Default: None return_convergence_delta (bool, optional): Indicates whether to return convergence delta or not. If `return_convergence_delta` is set to True convergence delta will be returned in a tuple following attributions. Default: False attribute_to_layer_input (bool, optional): Indicates whether to compute the attribution with respect to the layer input or output. If `attribute_to_layer_input` is set to True then the attributions will be computed with respect to layer input, otherwise it will be computed with respect to layer output. Note that currently it is assumed that either the input or the output of internal layer, depending on whether we attribute to the input or output, is a single tensor. Support for multiple tensors will be added later. Default: False custom_attribution_func (callable, optional): A custom function for computing final attribution scores. This function can take at least one and at most three arguments with the following signature: - custom_attribution_func(multipliers) - custom_attribution_func(multipliers, inputs) - custom_attribution_func(multipliers, inputs, baselines) In case this function is not provided, we use the default logic defined as: multipliers * (inputs - baselines) It is assumed that all input arguments, `multipliers`, `inputs` and `baselines` are provided in tuples of same length. `custom_attribution_func` returns a tuple of attribution tensors that have the same length as the `inputs`. Default: None Returns: **attributions** or 2-element tuple of **attributions**, **delta**: - **attributions** (*tensor* or tuple of *tensors*): Attribution score computed based on DeepLift's rescale rule with respect to layer's inputs or outputs. Attributions will always be the same size as the provided layer's inputs or outputs, depending on whether we attribute to the inputs or outputs of the layer. If the layer input / output is a single tensor, then just a tensor is returned; if the layer input / output has multiple tensors, then a corresponding tuple of tensors is returned. - **delta** (*tensor*, returned if return_convergence_delta=True): This is computed using the property that the total sum of forward_func(inputs) - forward_func(baselines) must equal the total sum of the attributions computed based on DeepLift's rescale rule. Delta is calculated per example, meaning that the number of elements in returned delta tensor is equal to the number of of examples in input. Note that the logic described for deltas is guaranteed when the default logic for attribution computations is used, meaning that the `custom_attribution_func=None`, otherwise it is not guaranteed and depends on the specifics of the `custom_attribution_func`. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. >>> net = ImageClassifier() >>> # creates an instance of LayerDeepLift to interpret target >>> # class 1 with respect to conv4 layer. >>> dl = LayerDeepLift(net, net.conv4) >>> input = torch.randn(1, 3, 32, 32, requires_grad=True) >>> # Computes deeplift attribution scores for conv4 layer and class 3. >>> attribution = dl.attribute(input, target=1) """ inputs = _format_input(inputs) baselines = _format_baseline(baselines, inputs) gradient_mask = apply_gradient_requirements(inputs) _validate_input(inputs, baselines) baselines = _tensorize_baseline(inputs, baselines) main_model_hooks = [] try: main_model_hooks = self._hook_main_model() self.model.apply(lambda mod: self._register_hooks( mod, attribute_to_layer_input=attribute_to_layer_input)) additional_forward_args = _format_additional_forward_args( additional_forward_args) expanded_target = _expand_target( target, 2, expansion_type=ExpansionTypes.repeat) wrapped_forward_func = self._construct_forward_func( self.model, (inputs, baselines), expanded_target, additional_forward_args, ) def chunk_output_fn( out: TensorOrTupleOfTensorsGeneric) -> Sequence: if isinstance(out, Tensor): return out.chunk(2) return tuple(out_sub.chunk(2) for out_sub in out) gradients, attrs = compute_layer_gradients_and_eval( wrapped_forward_func, self.layer, inputs, attribute_to_layer_input=attribute_to_layer_input, output_fn=lambda out: chunk_output_fn(out), ) attr_inputs = tuple(map(lambda attr: attr[0], attrs)) attr_baselines = tuple(map(lambda attr: attr[1], attrs)) gradients = tuple(map(lambda grad: grad[0], gradients)) if custom_attribution_func is None: if self.multiplies_by_inputs: attributions = tuple( (input - baseline) * gradient for input, baseline, gradient in zip( attr_inputs, attr_baselines, gradients)) else: attributions = gradients else: attributions = _call_custom_attribution_func( custom_attribution_func, gradients, attr_inputs, attr_baselines) finally: # remove hooks from all activations self._remove_hooks(main_model_hooks) undo_gradient_requirements(inputs, gradient_mask) return _compute_conv_delta_and_format_attrs( self, return_convergence_delta, attributions, baselines, inputs, additional_forward_args, target, cast(Union[Literal[True], Literal[False]], len(attributions) > 1), )
def _next_infidelity_tensors( current_n_perturb_samples: int, ) -> Union[Tuple[Tensor], Tuple[Tensor, Tensor, Tensor]]: perturbations, inputs_perturbed = _generate_perturbations( current_n_perturb_samples ) perturbations = _format_tensor_into_tuples(perturbations) inputs_perturbed = _format_tensor_into_tuples(inputs_perturbed) _validate_inputs_and_perturbations( cast(Tuple[Tensor, ...], inputs), cast(Tuple[Tensor, ...], inputs_perturbed), cast(Tuple[Tensor, ...], perturbations), ) targets_expanded = _expand_target( target, current_n_perturb_samples, expansion_type=ExpansionTypes.repeat_interleave, ) additional_forward_args_expanded = _expand_additional_forward_args( additional_forward_args, current_n_perturb_samples, expansion_type=ExpansionTypes.repeat_interleave, ) inputs_perturbed_fwd = _run_forward( forward_func, inputs_perturbed, targets_expanded, additional_forward_args_expanded, ) inputs_fwd = _run_forward(forward_func, inputs, target, additional_forward_args) inputs_fwd = torch.repeat_interleave( inputs_fwd, current_n_perturb_samples, dim=0 ) perturbed_fwd_diffs = inputs_fwd - inputs_perturbed_fwd attributions_expanded = tuple( torch.repeat_interleave(attribution, current_n_perturb_samples, dim=0) for attribution in attributions ) attributions_times_perturb = tuple( (attribution_expanded * perturbation).view(attribution_expanded.size(0), -1) for attribution_expanded, perturbation in zip( attributions_expanded, perturbations ) ) attr_times_perturb_sums = sum( torch.sum(attribution_times_perturb, dim=1) for attribution_times_perturb in attributions_times_perturb ) attr_times_perturb_sums = cast(Tensor, attr_times_perturb_sums) # reshape as Tensor(bsz, current_n_perturb_samples) attr_times_perturb_sums = attr_times_perturb_sums.view(bsz, -1) perturbed_fwd_diffs = perturbed_fwd_diffs.view(bsz, -1) if normalize: # in order to normalize, we have to aggregate the following tensors # to calculate MSE in its polynomial expansion: # (a-b)^2 = a^2 - 2ab + b^2 return ( attr_times_perturb_sums.pow(2).sum(-1), (attr_times_perturb_sums * perturbed_fwd_diffs).sum(-1), perturbed_fwd_diffs.pow(2).sum(-1), ) else: # returns (a-b)^2 if no need to normalize return ((attr_times_perturb_sums - perturbed_fwd_diffs).pow(2).sum(-1),)
def attribute( # type: ignore self, inputs: TensorOrTupleOfTensorsGeneric, baselines: BaselineType = None, target: TargetType = None, additional_forward_args: Any = None, return_convergence_delta: bool = False, custom_attribution_func: Union[None, Callable[..., Tuple[Tensor, ...]]] = None, ) -> Union[TensorOrTupleOfTensorsGeneric, Tuple[ TensorOrTupleOfTensorsGeneric, Tensor]]: r""" Args: inputs (tensor or tuple of tensors): Input for which 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 samples that are compared with the inputs. In order to assign attribution scores DeepLift computes the differences between the inputs/outputs and corresponding references. Baselines can be provided as: - a single tensor, if inputs is a single tensor, with exactly the same dimensions as inputs or the first dimension is one and the remaining dimensions match with 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. Note that attributions are not computed with respect to these arguments. Default: None return_convergence_delta (bool, optional): Indicates whether to return convergence delta or not. If `return_convergence_delta` is set to True convergence delta will be returned in a tuple following attributions. Default: False custom_attribution_func (callable, optional): A custom function for computing final attribution scores. This function can take at least one and at most three arguments with the following signature: - custom_attribution_func(multipliers) - custom_attribution_func(multipliers, inputs) - custom_attribution_func(multipliers, inputs, baselines) In case this function is not provided, we use the default logic defined as: multipliers * (inputs - baselines) It is assumed that all input arguments, `multipliers`, `inputs` and `baselines` are provided in tuples of same length. `custom_attribution_func` returns a tuple of attribution tensors that have the same length as the `inputs`. Default: None Returns: **attributions** or 2-element tuple of **attributions**, **delta**: - **attributions** (*tensor* or tuple of *tensors*): Attribution score computed based on DeepLift rescale rule 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. - **delta** (*tensor*, returned if return_convergence_delta=True): This is computed using the property that the total sum of forward_func(inputs) - forward_func(baselines) must equal the total sum of the attributions computed based on DeepLift's rescale rule. Delta is calculated per example, meaning that the number of elements in returned delta tensor is equal to the number of of examples in input. Note that the logic described for deltas is guaranteed when the default logic for attribution computations is used, meaning that the `custom_attribution_func=None`, otherwise it is not guaranteed and depends on the specifics of the `custom_attribution_func`. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. >>> net = ImageClassifier() >>> dl = DeepLift(net) >>> input = torch.randn(2, 3, 32, 32, requires_grad=True) >>> # Computes deeplift attribution scores for class 3. >>> attribution = dl.attribute(input, target=3) """ # Keeps track whether original input is a tuple or not before # converting it into a tuple. is_inputs_tuple = _is_tuple(inputs) inputs = _format_tensor_into_tuples(inputs) baselines = _format_baseline(baselines, inputs) gradient_mask = apply_gradient_requirements(inputs) _validate_input(inputs, baselines) # set hooks for baselines warnings.warn( """Setting forward, backward hooks and attributes on non-linear activations. The hooks and attributes will be removed after the attribution is finished""") baselines = _tensorize_baseline(inputs, baselines) main_model_hooks = [] try: main_model_hooks = self._hook_main_model() self.model.apply(self._register_hooks) additional_forward_args = _format_additional_forward_args( additional_forward_args) expanded_target = _expand_target( target, 2, expansion_type=ExpansionTypes.repeat) wrapped_forward_func = self._construct_forward_func( self.model, (inputs, baselines), expanded_target, additional_forward_args, ) gradients = self.gradient_func(wrapped_forward_func, inputs) if custom_attribution_func is None: if self.multiplies_by_inputs: attributions = tuple((input - baseline) * gradient for input, baseline, gradient in zip( inputs, baselines, gradients)) else: attributions = gradients else: attributions = _call_custom_attribution_func( custom_attribution_func, gradients, inputs, baselines) finally: # Even if any error is raised, remove all hooks before raising self._remove_hooks(main_model_hooks) undo_gradient_requirements(inputs, gradient_mask) return _compute_conv_delta_and_format_attrs( self, return_convergence_delta, attributions, baselines, inputs, additional_forward_args, target, is_inputs_tuple, )
def _attribute( self, inputs: Tuple[Tensor, ...], neuron_selector: Union[int, Tuple[int, ...], Callable], baselines: Tuple[Union[Tensor, int, float], ...], target: TargetType = None, additional_forward_args: Any = None, n_steps: int = 50, method: str = "riemann_trapezoid", attribute_to_neuron_input: bool = False, step_sizes_and_alphas: Union[None, Tuple[List[float], List[float]]] = None, ) -> Tuple[Tensor, ...]: num_examples = inputs[0].shape[0] total_batch = num_examples * n_steps if step_sizes_and_alphas is None: # retrieve step size and scaling factor for specified approximation method step_sizes_func, alphas_func = approximation_parameters(method) step_sizes, alphas = step_sizes_func(n_steps), alphas_func(n_steps) else: step_sizes, alphas = step_sizes_and_alphas # Compute scaled inputs from baseline to final input. scaled_features_tpl = tuple( torch.cat( [baseline + alpha * (input - baseline) for alpha in alphas], dim=0).requires_grad_() for input, baseline in zip(inputs, baselines)) additional_forward_args = _format_additional_forward_args( additional_forward_args) # apply number of steps to additional forward args # currently, number of steps is applied only to additional forward arguments # that are nd-tensors. It is assumed that the first dimension is # the number of batches. # dim -> (#examples * #steps x additional_forward_args[0].shape[1:], ...) input_additional_args = (_expand_additional_forward_args( additional_forward_args, n_steps) if additional_forward_args is not None else None) expanded_target = _expand_target(target, n_steps) # Conductance Gradients - Returns gradient of output with respect to # hidden layer and hidden layer evaluated at each input. layer_gradients, layer_eval, input_grads = compute_layer_gradients_and_eval( forward_fn=self.forward_func, layer=self.layer, inputs=scaled_features_tpl, target_ind=expanded_target, additional_forward_args=input_additional_args, gradient_neuron_selector=neuron_selector, device_ids=self.device_ids, attribute_to_layer_input=attribute_to_neuron_input, ) mid_grads = _verify_select_neuron(layer_gradients, neuron_selector) scaled_input_gradients = tuple( input_grad * mid_grads.reshape((total_batch, ) + (1, ) * (len(input_grad.shape) - 1)) for input_grad in input_grads) # Mutliplies by appropriate step size. scaled_grads = tuple( scaled_input_gradient.contiguous().view(n_steps, -1) * torch.tensor(step_sizes).view(n_steps, 1).to( scaled_input_gradient.device) for scaled_input_gradient in scaled_input_gradients) # Aggregates across all steps for each tensor in the input tuple total_grads = tuple( _reshape_and_sum(scaled_grad, n_steps, num_examples, input_grad.shape[1:]) for (scaled_grad, input_grad) in zip(scaled_grads, input_grads)) if self.multiplies_by_inputs: # computes attribution for each tensor in input tuple # attributions has the same dimensionality as inputs attributions = tuple(total_grad * (input - baseline) for total_grad, input, baseline in zip( total_grads, inputs, baselines)) else: attributions = total_grads return attributions
def _attribute( self, inputs: Tuple[Tensor, ...], baselines: Tuple[Union[Tensor, int, float], ...], target: TargetType = None, additional_forward_args: Any = None, n_steps: int = 50, method: str = "gausslegendre", attribute_to_layer_input: bool = False, step_sizes_and_alphas: Union[None, Tuple[List[float], List[float]]] = None, ) -> Union[Tensor, Tuple[Tensor, ...]]: if step_sizes_and_alphas is None: # retrieve step size and scaling factor for specified approximation method step_sizes_func, alphas_func = approximation_parameters(method) step_sizes, alphas = step_sizes_func(n_steps), alphas_func(n_steps) else: step_sizes, alphas = step_sizes_and_alphas # Compute scaled inputs from baseline to final input. scaled_features_tpl = tuple( torch.cat( [baseline + alpha * (input - baseline) for alpha in alphas], dim=0).requires_grad_() for input, baseline in zip(inputs, baselines)) additional_forward_args = _format_additional_forward_args( additional_forward_args) # apply number of steps to additional forward args # currently, number of steps is applied only to additional forward arguments # that are nd-tensors. It is assumed that the first dimension is # the number of batches. # dim -> (bsz * #steps x additional_forward_args[0].shape[1:], ...) input_additional_args = (_expand_additional_forward_args( additional_forward_args, n_steps) if additional_forward_args is not None else None) expanded_target = _expand_target(target, n_steps) # Returns gradient of output with respect to hidden layer. layer_gradients, _ = compute_layer_gradients_and_eval( forward_fn=self.forward_func, layer=self.layer, inputs=scaled_features_tpl, target_ind=expanded_target, additional_forward_args=input_additional_args, device_ids=self.device_ids, attribute_to_layer_input=attribute_to_layer_input, ) # flattening grads so that we can multiply it with step-size # calling contiguous to avoid `memory whole` problems scaled_grads = tuple( layer_grad.contiguous().view(n_steps, -1) * torch.tensor(step_sizes).view(n_steps, 1).to(layer_grad.device) for layer_grad in layer_gradients) # aggregates across all steps for each tensor in the input tuple attrs = tuple( _reshape_and_sum(scaled_grad, n_steps, inputs[0].shape[0], layer_grad.shape[1:]) for scaled_grad, layer_grad in zip(scaled_grads, layer_gradients)) return _format_output(len(attrs) > 1, attrs)
def _ablation_generator( self, i, inputs, additional_args, target, baselines, input_mask, perturbations_per_eval, **kwargs ): """ This method is a generator which yields each perturbation to be evaluated including inputs, additional_forward_args, targets, and 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 _attribute( self, inputs: Tuple[Tensor, ...], baselines: Tuple[Union[Tensor, int, float], ...], target: TargetType = None, additional_forward_args: Any = None, n_steps: int = 50, method: str = "gausslegendre", step_sizes_and_alphas: Union[None, Tuple[List[float], List[float]]] = None, ) -> Tuple[Tensor, ...]: if step_sizes_and_alphas is None: # retrieve step size and scaling factor for specified # approximation method step_sizes_func, alphas_func = approximation_parameters(method) step_sizes, alphas = step_sizes_func(n_steps), alphas_func(n_steps) else: step_sizes, alphas = step_sizes_and_alphas # scale features and compute gradients. (batch size is abbreviated as bsz) # scaled_features' dim -> (bsz * #steps x inputs[0].shape[1:], ...) scaled_features_tpl = tuple( torch.cat( [baseline + alpha * (input - baseline) for alpha in alphas], dim=0 ).requires_grad_() for input, baseline in zip(inputs, baselines) ) additional_forward_args = _format_additional_forward_args( additional_forward_args ) # apply number of steps to additional forward args # currently, number of steps is applied only to additional forward arguments # that are nd-tensors. It is assumed that the first dimension is # the number of batches. # dim -> (bsz * #steps x additional_forward_args[0].shape[1:], ...) input_additional_args = ( _expand_additional_forward_args(additional_forward_args, n_steps) if additional_forward_args is not None else None ) expanded_target = _expand_target(target, n_steps) # grads: dim -> (bsz * #steps x inputs[0].shape[1:], ...) grads = self.gradient_func( forward_fn=self.forward_func, inputs=scaled_features_tpl, target_ind=expanded_target, additional_forward_args=input_additional_args, ) # flattening grads so that we can multilpy it with step-size # calling contiguous to avoid `memory whole` problems scaled_grads = [ grad.contiguous().view(n_steps, -1) * torch.tensor(step_sizes).view(n_steps, 1).to(grad.device) for grad in grads ] # aggregates across all steps for each tensor in the input tuple # total_grads has the same dimensionality as inputs total_grads = tuple( _reshape_and_sum( scaled_grad, n_steps, grad.shape[0] // n_steps, grad.shape[1:] ) for (scaled_grad, grad) in zip(scaled_grads, grads) ) # computes attribution for each tensor in input tuple # attributions has the same dimensionality as inputs if not self.multiplies_by_inputs: attributions = total_grads else: attributions = tuple( total_grad * (input - baseline) for total_grad, input, baseline in zip(total_grads, inputs, baselines) ) return attributions
def _perturbation_generator( self, inputs: Tuple[Tensor, ...], additional_args: Any, target: TargetType, baselines: Tuple[Tensor, ...], input_masks: TensorOrTupleOfTensorsGeneric, feature_permutation: Sequence[int], perturbations_per_eval: int, ) -> Iterable[Tuple[Tuple[Tensor, ...], Any, TargetType, Tuple[Tensor, ...]]]: """ This method is a generator which yields each perturbation to be evaluated including inputs, additional_forward_args, targets, and mask. """ # current_tensors starts at baselines and includes each additional feature as # added based on the permutation order. current_tensors = baselines current_tensors_list = [] current_mask_list = [] # Compute repeated additional args and targets 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) for i in range(len(feature_permutation)): current_tensors = tuple( current * (~(mask == feature_permutation[i])).to(current.dtype) + input * (mask == feature_permutation[i]).to(input.dtype) for input, current, mask in zip(inputs, current_tensors, input_masks) ) current_tensors_list.append(current_tensors) current_mask_list.append( tuple(mask == feature_permutation[i] for mask in input_masks) ) if len(current_tensors_list) == perturbations_per_eval: combined_inputs = tuple( torch.cat(aligned_tensors, dim=0) for aligned_tensors in zip(*current_tensors_list) ) combined_masks = tuple( torch.stack(aligned_masks, dim=0) for aligned_masks in zip(*current_mask_list) ) yield ( combined_inputs, additional_args_repeated, target_repeated, combined_masks, ) current_tensors_list = [] current_mask_list = [] # Create batch with remaining evaluations, may not be a complete batch # (= perturbations_per_eval) if len(current_tensors_list) != 0: additional_args_repeated = ( _expand_additional_forward_args( additional_args, len(current_tensors_list) ) if additional_args is not None else None ) target_repeated = _expand_target(target, len(current_tensors_list)) combined_inputs = tuple( torch.cat(aligned_tensors, dim=0) for aligned_tensors in zip(*current_tensors_list) ) combined_masks = tuple( torch.stack(aligned_masks, dim=0) for aligned_masks in zip(*current_mask_list) ) yield ( combined_inputs, additional_args_repeated, target_repeated, combined_masks, )
def _attribute( self, inputs: Tuple[Tensor, ...], baselines: Tuple[Union[Tensor, int, float], ...], target: TargetType = None, additional_forward_args: Any = None, n_steps: int = 50, method: str = "gausslegendre", attribute_to_layer_input: bool = False, step_sizes_and_alphas: Union[None, Tuple[List[float], List[float]]] = None, ) -> Union[Tensor, Tuple[Tensor, ...]]: num_examples = inputs[0].shape[0] if step_sizes_and_alphas is None: # Retrieve scaling factors for specified approximation method step_sizes_func, alphas_func = approximation_parameters(method) alphas = alphas_func(n_steps + 1) else: _, alphas = step_sizes_and_alphas # Compute scaled inputs from baseline to final input. scaled_features_tpl = tuple( torch.cat( [baseline + alpha * (input - baseline) for alpha in alphas], dim=0 ).requires_grad_() for input, baseline in zip(inputs, baselines) ) additional_forward_args = _format_additional_forward_args( additional_forward_args ) # apply number of steps to additional forward args # currently, number of steps is applied only to additional forward arguments # that are nd-tensors. It is assumed that the first dimension is # the number of batches. # dim -> (#examples * #steps x additional_forward_args[0].shape[1:], ...) input_additional_args = ( _expand_additional_forward_args(additional_forward_args, n_steps + 1) if additional_forward_args is not None else None ) expanded_target = _expand_target(target, n_steps + 1) # Conductance Gradients - Returns gradient of output with respect to # hidden layer and hidden layer evaluated at each input. (layer_gradients, layer_evals,) = compute_layer_gradients_and_eval( forward_fn=self.forward_func, layer=self.layer, inputs=scaled_features_tpl, additional_forward_args=input_additional_args, target_ind=expanded_target, device_ids=self.device_ids, attribute_to_layer_input=attribute_to_layer_input, ) # Compute differences between consecutive evaluations of layer_eval. # This approximates the total input gradient of each step multiplied # by the step size. grad_diffs = tuple( layer_eval[num_examples:] - layer_eval[:-num_examples] for layer_eval in layer_evals ) # Element-wise multiply gradient of output with respect to hidden layer # and summed gradients with respect to input (chain rule) and sum # across stepped inputs. attributions = tuple( _reshape_and_sum( grad_diff * layer_gradient[:-num_examples], n_steps, num_examples, layer_eval.shape[1:], ) for layer_gradient, layer_eval, grad_diff in zip( layer_gradients, layer_evals, grad_diffs ) ) return _format_output(len(attributions) > 1, attributions)