def attribute( self, inputs: TensorOrTupleOfTensorsGeneric, target: TargetType = None, additional_forward_args: Any = None, ) -> TensorOrTupleOfTensorsGeneric: r""" Computes attribution by overriding relu gradients. Based on constructor flag use_relu_grad_output, performs either GuidedBackpropagation if False and Deconvolution if True. This class is the parent class of both these methods, more information on usage can be found in the docstrings for each implementing class. """ # Keeps track whether original input is a tuple or not before # converting it into a tuple. is_inputs_tuple = _is_tuple(inputs) inputs = _format_input(inputs) gradient_mask = apply_gradient_requirements(inputs) # set hooks for overriding ReLU gradients warnings.warn( "Setting backward hooks on ReLU activations." "The hooks will be removed after the attribution is finished") try: self.model.apply(self._register_hooks) gradients = self.gradient_func(self.forward_func, inputs, target, additional_forward_args) finally: self._remove_hooks() undo_gradient_requirements(inputs, gradient_mask) return _format_output(is_inputs_tuple, gradients)
def test_apply_gradient_reqs(self) -> None: initial_grads = [False, True, False] test_tensor = torch.tensor([[6.0]], requires_grad=True) test_tensor.grad = torch.tensor([[7.0]]) test_tensor_tuple = (torch.tensor([[5.0]]), test_tensor, torch.tensor([[7.0]])) out_mask = apply_gradient_requirements(test_tensor_tuple) for i in range(len(test_tensor_tuple)): self.assertTrue(test_tensor_tuple[i].requires_grad) self.assertEqual(out_mask[i], initial_grads[i])
def attribute(self, inputs: Union[Tensor, Tuple[Tensor, ...]], target: TargetType = None, additional_forward_args: Any = None, attribute_to_layer_input: bool = False, relu_attributions: bool = False) -> Union[Tensor, Tuple[Tensor, ...]]: inputs = _format_input(inputs) additional_forward_args = _format_additional_forward_args( additional_forward_args ) gradient_mask = apply_gradient_requirements(inputs) # 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( self.forward_func, self.layer, inputs, target, additional_forward_args, device_ids=self.device_ids, attribute_to_layer_input=attribute_to_layer_input, ) undo_gradient_requirements(inputs, gradient_mask) summed_grads = tuple( torch.mean( layer_grad, dim=0, keepdim=True, ) for layer_grad in layer_gradients ) scaled_acts = tuple( torch.sum(summed_grad * layer_eval, dim=1, keepdim=True) for summed_grad, layer_eval in zip(summed_grads, layer_evals) ) if relu_attributions: scaled_acts = tuple(F.relu(scaled_act) for scaled_act in scaled_acts) return _format_output(len(scaled_acts) > 1, scaled_acts)
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 attribute( self, inputs: TensorOrTupleOfTensorsGeneric, target: TargetType = None, additional_forward_args: Any = None, ) -> TensorOrTupleOfTensorsGeneric: 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. 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 Returns: *tensor* or tuple of *tensors* of **attributions**: - **attributions** (*tensor* or tuple of *tensors*): The input x gradient 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. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. >>> net = ImageClassifier() >>> # Generating random input with size 2x3x3x32 >>> input = torch.randn(2, 3, 32, 32, requires_grad=True) >>> # Defining InputXGradient interpreter >>> input_x_gradient = InputXGradient(net) >>> # Computes inputXgradient for class 4. >>> attribution = input_x_gradient.attribute(input, target=4) """ # Keeps track whether original input is a tuple or not before # converting it into a tuple. is_inputs_tuple = _is_tuple(inputs) inputs = _format_input(inputs) gradient_mask = apply_gradient_requirements(inputs) gradients = self.gradient_func(self.forward_func, inputs, target, additional_forward_args) attributions = tuple(input * gradient for input, gradient in zip(inputs, gradients)) undo_gradient_requirements(inputs, gradient_mask) return _format_output(is_inputs_tuple, attributions)
def attribute( self, inputs: TensorOrTupleOfTensorsGeneric, neuron_selector: Union[int, Tuple[Union[int, slice], ...], Callable], additional_forward_args: Any = None, attribute_to_neuron_input: bool = False, ) -> TensorOrTupleOfTensorsGeneric: r""" Args: inputs (tensor or tuple of tensors): Input for which neuron gradients 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, and if multiple input tensors are provided, the examples must be aligned appropriately. neuron_selector (int, callable, or tuple of ints or slices): Selector for neuron in given layer for which attribution is desired. Neuron selector can be provided as: - a single integer, if the layer output is 2D. This integer selects the appropriate neuron column in the layer input or output - a tuple of integers or slice objects. Length of this tuple must be one less than the number of dimensions in the input / output of the given layer (since dimension 0 corresponds to number of examples). The elements of the tuple can be either integers or slice objects (slice object allows indexing a range of neurons rather individual ones). If any of the tuple elements is a slice object, the indexed output tensor is used for attribution. Note that specifying a slice of a tensor would amount to computing the attribution of the sum of the specified neurons, and not the individual neurons independantly. - a callable, which should take the target layer as input (single tensor or tuple if multiple tensors are in layer) and return a neuron or aggregate of the layer's neurons for attribution. For example, this function could return the sum of the neurons in the layer or sum of neurons with activations in a particular range. It is expected that this function returns either a tensor with one element or a 1D tensor with length equal to batch_size (one scalar per input 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. Note that attributions are not computed with respect to these arguments. Default: None attribute_to_neuron_input (bool, optional): Indicates whether to compute the attributions with respect to the neuron input or output. If `attribute_to_neuron_input` is set to True then the attributions will be computed with respect to neuron's inputs, otherwise it will be computed with respect to neuron's outputs. Note that currently it is assumed that either the input or the output of internal neurons, depending on whether we attribute to the input or output, is a single tensor. Support for multiple tensors will be added later. Default: False Returns: *tensor* or tuple of *tensors* of **attributions**: - **attributions** (*tensor* or tuple of *tensors*): Gradients of particular neuron 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. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. >>> # It contains an attribute conv1, which is an instance of nn.conv2d, >>> # and the output of this layer has dimensions Nx12x32x32. >>> net = ImageClassifier() >>> neuron_ig = NeuronGradient(net, net.conv1) >>> input = torch.randn(2, 3, 32, 32, requires_grad=True) >>> # To compute neuron attribution, we need to provide the neuron >>> # index for which attribution is desired. Since the layer output >>> # is Nx12x32x32, we need a tuple in the form (0..11,0..31,0..31) >>> # which indexes a particular neuron in the layer output. >>> # For this example, we choose the index (4,1,2). >>> # Computes neuron gradient for neuron with >>> # index (4,1,2). >>> attribution = neuron_ig.attribute(input, (4,1,2)) """ is_inputs_tuple = _is_tuple(inputs) inputs = _format_tensor_into_tuples(inputs) additional_forward_args = _format_additional_forward_args( additional_forward_args ) gradient_mask = apply_gradient_requirements(inputs) _, input_grads = _forward_layer_eval_with_neuron_grads( self.forward_func, inputs, self.layer, additional_forward_args, gradient_neuron_selector=neuron_selector, device_ids=self.device_ids, attribute_to_layer_input=attribute_to_neuron_input, ) undo_gradient_requirements(inputs, gradient_mask) return _format_output(is_inputs_tuple, input_grads)
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: TensorOrTupleOfTensorsGeneric, target: TargetType = None, additional_forward_args: Any = None, return_convergence_delta: bool = False, attribute_to_layer_input: bool = False, verbose: bool = False, ) -> Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[ Tensor, ...]]], Tuple[Union[Tensor, Tuple[Tensor, ...], List[Union[ Tensor, Tuple[Tensor, ...]]]], Union[Tensor, List[Tensor]], ], ]: r""" Args: inputs (tensor or tuple of tensors): Input for which relevance is propagated. 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, and if multiple input tensors are provided, the examples must be aligned appropriately. 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 (tuple, 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. verbose (bool, optional): Indicates whether information on application of rules is printed during propagation. Default: False Returns: *tensor* or tuple of *tensors* of **attributions** or 2-element tuple of **attributions**, **delta** or lists of **attributions** and **delta**: - **attributions** (*tensor* or tuple of *tensors*): The propagated relevance values 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. The sum of attributions is one and not corresponding to the prediction score as in other implementations. If attributions for all layers are returned (layer=None) a list of tensors or tuples of tensors is returned with entries for each layer. - **delta** (*tensor* or list of *tensors* returned if return_convergence_delta=True): 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. If attributions for all layers are returned (layer=None) a list of tensors is returned with entries for each layer. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. It has one >>> # Conv2D and a ReLU layer. >>> net = ImageClassifier() >>> lrp = LRP(net, net.conv1) >>> input = torch.randn(3, 3, 32, 32) >>> # Attribution size matches input size: 3x3x32x32 >>> attribution = lrp.attribute(input, target=5) """ self.verbose = verbose self._original_state_dict = self.model.state_dict() self.layers = [] self._get_layers(self.model) self._check_and_attach_rules() self.attribute_to_layer_input = attribute_to_layer_input self.backward_handles = [] self.forward_handles = [] inputs = _format_tensor_into_tuples(inputs) gradient_mask = apply_gradient_requirements(inputs) try: # 1. Forward pass output = self._compute_output_and_change_weights( inputs, target, additional_forward_args) self._register_forward_hooks() # 2. Forward pass + backward pass _ = compute_gradients(self._forward_fn_wrapper, inputs, target, additional_forward_args) relevances = self._get_output_relevance(output) finally: self._restore_model() undo_gradient_requirements(inputs, gradient_mask) if return_convergence_delta: delta: Union[Tensor, List[Tensor]] if isinstance(self.layer, list): delta = [] for relevance_layer in relevances: delta.append( self.compute_convergence_delta(relevance_layer, output)) else: delta = self.compute_convergence_delta( cast(Tuple[Tensor, ...], relevances), output) return relevances, delta # type: ignore else: return relevances # type: ignore
def perturb( self, inputs: TensorOrTupleOfTensorsGeneric, epsilon: float, target: Any, additional_forward_args: Any = None, targeted: bool = False, ) -> TensorOrTupleOfTensorsGeneric: r""" This method computes and returns the perturbed input for each input tensor. It supports both targeted and non-targeted attacks. Args: inputs (tensor or tuple of tensors): Input for which adversarial attack is computed. It can be provided as a single tensor or a tuple of multiple tensors. If multiple input tensors are provided, the batch sizes must be aligned accross all tensors. epsilon (float): Step size of perturbation. target (any): True labels of inputs if non-targeted attack is desired. Target class of inputs if targeted attack is desired. Target will be passed to the loss function to compute loss, so the type needs to match the argument type of the loss function. If using the default negative log as loss function, labels should be of type int, tuple, tensor or list. For general 2D outputs, labels 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 label for the corresponding example. For outputs with > 2 dimensions, labels can be either: - A single tuple, which contains #output_dims - 1 elements. This label 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 label 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. These arguments are provided to forward_func in order following the arguments in inputs. Default: None. targeted (bool, optional): If attack should be targeted. Default: False. Returns: - **perturbed inputs** (*tensor* or tuple of *tensors*): Perturbed input for each input tensor. The perturbed inputs have the same shape and dimensionality as the inputs. 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. """ is_inputs_tuple = _is_tuple(inputs) inputs: Tuple[Tensor, ...] = _format_input(inputs) gradient_mask = apply_gradient_requirements(inputs) def _forward_with_loss() -> Tensor: additional_inputs = _format_additional_forward_args(additional_forward_args) outputs = self.forward_func( # type: ignore *(*inputs, *additional_inputs) # type: ignore if additional_inputs is not None else inputs ) if self.loss_func is not None: return self.loss_func(outputs, target) else: loss = -torch.log(outputs) return _select_targets(loss, target) grads = compute_gradients(_forward_with_loss, inputs) undo_gradient_requirements(inputs, gradient_mask) perturbed_inputs = self._perturb(inputs, grads, epsilon, targeted) perturbed_inputs = tuple( self.bound(perturbed_inputs[i]) for i in range(len(perturbed_inputs)) ) return _format_output(is_inputs_tuple, perturbed_inputs)
def attribute( self, inputs: TensorOrTupleOfTensorsGeneric, target: TargetType = None, additional_forward_args: Any = None, return_convergence_delta: bool = False, verbose: bool = False, ) -> Union[ TensorOrTupleOfTensorsGeneric, Tuple[TensorOrTupleOfTensorsGeneric, Tensor] ]: r""" Args: inputs (tensor or tuple of tensors): Input for which relevance is propagated. 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, and if multiple input tensors are provided, the examples must be aligned appropriately. 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 (tuple, 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 verbose (bool, optional): Indicates whether information on application of rules is printed during propagation. Returns: *tensor* or tuple of *tensors* of **attributions** or 2-element tuple of **attributions**, **delta**:: - **attributions** (*tensor* or tuple of *tensors*): The propagated relevance values with respect to each input feature. The values are normalized by the output score value (sum(relevance)=1). To obtain values comparable to other methods or implementations these values need to be multiplied by the output score. 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. The sum of attributions is one and not corresponding to the prediction score as in other implementations. - **delta** (*tensor*, returned if return_convergence_delta=True): Delta is calculated per example, meaning that the number of elements in returned delta tensor is equal to the number of of examples in the inputs. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. It has one >>> # Conv2D and a ReLU layer. >>> net = ImageClassifier() >>> lrp = LRP(net) >>> input = torch.randn(3, 3, 32, 32) >>> # Attribution size matches input size: 3x3x32x32 >>> attribution = lrp.attribute(input, target=5) """ self.verbose = verbose self._original_state_dict = self.model.state_dict() self.layers: List[Module] = [] self._get_layers(self.model) self._check_and_attach_rules() self.backward_handles: List[RemovableHandle] = [] self.forward_handles: List[RemovableHandle] = [] is_inputs_tuple = _is_tuple(inputs) inputs = _format_input(inputs) gradient_mask = apply_gradient_requirements(inputs) try: # 1. Forward pass: Change weights of layers according to selected rules. output = self._compute_output_and_change_weights( inputs, target, additional_forward_args ) # 2. Forward pass + backward pass: Register hooks to configure relevance # propagation and execute back-propagation. self._register_forward_hooks() normalized_relevances = self.gradient_func( self._forward_fn_wrapper, inputs, target, additional_forward_args ) relevances = tuple( normalized_relevance * output.reshape((-1,) + (1,) * (normalized_relevance.dim() - 1)) for normalized_relevance in normalized_relevances ) finally: self._restore_model() undo_gradient_requirements(inputs, gradient_mask) if return_convergence_delta: return ( _format_output(is_inputs_tuple, relevances), self.compute_convergence_delta(relevances, output), ) else: return _format_output(is_inputs_tuple, relevances) # type: ignore
def attribute( self, inputs: Union[Tensor, Tuple[Tensor, ...]], target: TargetType = None, additional_forward_args: Any = None, attribute_to_layer_input: bool = False, relu_attributions: bool = False, ) -> Union[Tensor, Tuple[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, and if multiple input tensors are provided, the examples must be aligned appropriately. 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 attribute_to_layer_input (bool, optional): Indicates whether to compute the attributions 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 the layer input, otherwise it will be computed with respect to layer output. Note that currently it is assumed that either the input or the outputs of internal layers, depending on whether we attribute to the input or output, are single tensors. Support for multiple tensors will be added later. Default: False relu_attributions (bool, optional): Indicates whether to apply a ReLU operation on the final attribution, returning only non-negative attributions. Setting this flag to True matches the original GradCAM algorithm, otherwise, by default, both positive and negative attributions are returned. Default: False Returns: *tensor* or tuple of *tensors* of **attributions**: - **attributions** (*tensor* or tuple of *tensors*): Attributions based on GradCAM method. Attributions will be the same size as the output of the given layer, except for dimension 2, which will be 1 due to summing over channels. Attributions are returned in a tuple if the layer inputs / outputs contain multiple tensors, otherwise a single tensor is returned. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. >>> # It contains a layer conv4, which is an instance of nn.conv2d, >>> # and the output of this layer has dimensions Nx50x8x8. >>> # It is the last convolution layer, which is the recommended >>> # use case for GradCAM. >>> net = ImageClassifier() >>> layer_gc = LayerGradCam(net, net.conv4) >>> input = torch.randn(2, 3, 32, 32, requires_grad=True) >>> # Computes layer GradCAM for class 3. >>> # attribution size matches layer output except for dimension >>> # 1, so dimensions of attr would be Nx1x8x8. >>> attr = layer_gc.attribute(input, 3) >>> # GradCAM attributions are often upsampled and viewed as a >>> # mask to the input, since the convolutional layer output >>> # spatially matches the original input image. >>> # This can be done with LayerAttribution's interpolate method. >>> upsampled_attr = LayerAttribution.interpolate(attr, (32, 32)) """ inputs = _format_input(inputs) additional_forward_args = _format_additional_forward_args( additional_forward_args) gradient_mask = apply_gradient_requirements(inputs) # 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( self.forward_func, self.layer, inputs, target, additional_forward_args, device_ids=self.device_ids, attribute_to_layer_input=attribute_to_layer_input, ) undo_gradient_requirements(inputs, gradient_mask) summed_grads = tuple( torch.mean( layer_grad, dim=tuple(x for x in range(2, len(layer_grad.shape))), keepdim=True, ) if len(layer_grad.shape) > 2 else layer_grad for layer_grad in layer_gradients) scaled_acts = tuple( torch.sum(summed_grad * layer_eval, dim=1, keepdim=True) for summed_grad, layer_eval in zip(summed_grads, layer_evals)) if relu_attributions: scaled_acts = tuple( F.relu(scaled_act) for scaled_act in scaled_acts) return _format_output(len(scaled_acts) > 1, scaled_acts)
def attribute( self, inputs: Union[Tensor, Tuple[Tensor, ...]], target: TargetType = None, additional_forward_args: Any = None, attribute_to_layer_input: bool = False, ) -> Union[Tensor, Tuple[Tensor, ...], List[Union[Tensor, Tuple[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, and if multiple input tensors are provided, the examples must be aligned appropriately. 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 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. Default: False Returns: *tensor* or tuple of *tensors* or *list* of **attributions**: - **attributions** (*tensor* or tuple of *tensors* or *list*): Product of gradient and activation for each neuron in given layer output. Attributions will always be the same size as the output of the given layer. Attributions are returned in a tuple if the layer inputs / outputs contain multiple tensors, otherwise a single tensor is returned. If multiple layers are provided, attributions are returned as a list, each element corresponding to the activations of the corresponding layer. Examples:: >>> # ImageClassifier takes a single input tensor of images Nx3x32x32, >>> # and returns an Nx10 tensor of class probabilities. >>> # It contains an attribute conv1, which is an instance of nn.conv2d, >>> # and the output of this layer has dimensions Nx12x32x32. >>> net = ImageClassifier() >>> layer_ga = LayerGradientXActivation(net, net.conv1) >>> input = torch.randn(2, 3, 32, 32, requires_grad=True) >>> # Computes layer activation x gradient for class 3. >>> # attribution size matches layer output, Nx12x32x32 >>> attribution = layer_ga.attribute(input, 3) """ inputs = _format_input(inputs) additional_forward_args = _format_additional_forward_args( additional_forward_args ) gradient_mask = apply_gradient_requirements(inputs) # 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( self.forward_func, self.layer, inputs, target, additional_forward_args, device_ids=self.device_ids, attribute_to_layer_input=attribute_to_layer_input, ) undo_gradient_requirements(inputs, gradient_mask) if isinstance(self.layer, Module): return _format_output( len(layer_evals) > 1, self.multiply_gradient_acts(layer_gradients, layer_evals), ) else: return [ _format_output( len(layer_evals[i]) > 1, self.multiply_gradient_acts(layer_gradients[i], layer_evals[i]), ) for i in range(len(self.layer)) ]
def attribute( self, inputs, baselines=None, target=None, n_steps=500, method="riemann_trapezoid", ): r""" Computes conductance using gradients along the path, applying riemann's method or gauss-legendre. The details of the approach can be found here: https://arxiv.org/abs/1805.12233 Args inputs: A single high dimensional input tensor, in which dimension 0 corresponds to number of examples. baselines: A single high dimensional baseline tensor, which has the same shape as the input target: Predicted class index. This is necessary only for classification use cases n_steps: The number of steps used by the approximation method method: Method for integral approximation, one of `riemann_right`, `riemann_middle`, `riemann_trapezoid` or `gausslegendre` Return attributions: Total conductance with respect to each neuron in output of given layer """ if baselines is None: baselines = 0 gradient_mask = apply_gradient_requirements((inputs, )) # 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) # compute scaled inputs from baseline to final input. scaled_features = torch.cat( [baselines + alpha * (inputs - baselines) for alpha in alphas], dim=0) # Conductance Gradients - Returns gradient of output with respect to # hidden layer, gradient of hidden layer with respect to input, # and number of hidden units. input_gradients, mid_layer_gradients, hidden_units = self._conductance_grads( self.forward_func, scaled_features, target) # Multiply gradient of hidden layer with respect to input by input - baseline scaled_input_gradients = torch.repeat_interleave(inputs - baselines, hidden_units, dim=0) scaled_input_gradients = input_gradients * scaled_input_gradients.repeat( *([len(alphas)] + [1] * (len(scaled_input_gradients.shape) - 1))) # Sum gradients for each input neuron in order to have total # for each hidden unit and reshape to match hidden layer shape summed_input_grads = torch.sum( scaled_input_gradients, tuple(range(1, len( scaled_input_gradients.shape)))).view_as(mid_layer_gradients) # Rescale gradients of hidden layer by by step size. scaled_grads = mid_layer_gradients.contiguous().view( n_steps, -1) * torch.tensor(step_sizes).view(n_steps, 1).to( mid_layer_gradients.device) undo_gradient_requirements((inputs, ), gradient_mask) # Element-wise mutliply gradient of output with respect to hidden layer # and summed gradients with respect to input (chain rule) and sum across # stepped inputs. return _reshape_and_sum( scaled_grads.view(mid_layer_gradients.shape) * summed_input_grads, n_steps, inputs.shape[0], mid_layer_gradients.shape[1:], )