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 compute_gradients(
forward_fn: Callable,
inputs: Union[Tensor, Tuple[Tensor, ...]],
target_ind: TargetType = None,
additional_forward_args: Any = None,
) -> Tuple[Tensor, ...]:
r"""
Computes gradients of the output with respect to inputs for an
arbitrary forward function.
Args:
forward_fn: forward function. This can be for example model's
forward function.
input: Input at which gradients are evaluated,
will be passed to forward_fn.
target_ind: Index of the target class for which gradients
must be computed (classification only).
additional_forward_args: Additional input arguments that forward
function requires. It takes an empty tuple (no additional
arguments) if no additional arguments are required
"""
with torch.autograd.set_grad_enabled(True):
# runs forward pass
outputs = _run_forward(forward_fn, inputs, target_ind, additional_forward_args)
assert outputs[0].numel() == 1, (
"Target not provided when necessary, cannot"
" take gradient with respect to multiple outputs."
)
# torch.unbind(forward_out) is a list of scalar tensor tuples and
# contains batch_size * #steps elements
grads = torch.autograd.grad(torch.unbind(outputs), inputs)
return grads
def forward_fn():
model_out = _run_forward(
forward_func, inputs, None, additional_forward_args
)
return _select_targets(
torch.cat((model_out[:, 0], model_out[:, 1])), target
)
def calculate_predicted_scores(
self, inputs, additional_forward_args, model
) -> Tuple[
List[OutputScore], Optional[List[Tuple[Tensor, ...]]], Tuple[Tensor, ...]
]:
# Check if inputs have cached baselines and transformed inputs
hashable_inputs = tuple(inputs)
if hashable_inputs in self.baseline_cache:
baselines_group = self.baseline_cache[hashable_inputs]
transformed_inputs = self.transformed_input_cache[hashable_inputs]
else:
# Initialize baselines
baseline_transforms_len = 1 # todo support multiple baselines
baselines: List[List[Optional[Tensor]]] = [
[None] * len(self.features) for _ in range(baseline_transforms_len)
]
transformed_inputs = list(inputs)
for feature_i, feature in enumerate(self.features):
transformed_inputs[feature_i] = self._transform(
feature.input_transforms, transformed_inputs[feature_i], True
)
for baseline_i in range(baseline_transforms_len):
if baseline_i > len(feature.baseline_transforms) - 1:
baselines[baseline_i][feature_i] = torch.zeros_like(
transformed_inputs[feature_i]
)
else:
baselines[baseline_i][feature_i] = self._transform(
[feature.baseline_transforms[baseline_i]],
transformed_inputs[feature_i],
True,
)
baselines = cast(List[List[Optional[Tensor]]], baselines)
baselines_group = [tuple(b) for b in baselines]
self.baseline_cache[hashable_inputs] = baselines_group
self.transformed_input_cache[hashable_inputs] = transformed_inputs
outputs = _run_forward(
model,
tuple(transformed_inputs),
additional_forward_args=additional_forward_args,
)
if self.score_func is not None:
outputs = self.score_func(outputs)
if outputs.nelement() == 1:
scores = outputs
predicted = scores.round().to(torch.int)
else:
scores, predicted = outputs.topk(min(4, outputs.shape[-1]))
scores = scores.cpu().squeeze(0)
predicted = predicted.cpu().squeeze(0)
predicted_scores = self._get_labels_from_scores(scores, predicted)
return predicted_scores, baselines_group, tuple(transformed_inputs)
def _compute_output_and_change_weights(
self,
inputs: Tuple[Tensor, ...],
target: TargetType,
additional_forward_args: Any,
) -> Tensor:
try:
self._register_weight_hooks()
output = _run_forward(self.model, inputs, target, additional_forward_args)
finally:
self._remove_forward_hooks()
# Register pre_hooks that pass the initial activations from before weight
# adjustments as inputs to the layers with adjusted weights. This procedure
# is important for graph generation in the 2nd forward pass.
self._register_pre_hooks()
return output
def layer_forward_func(*args):
layer_length = args[-1]
layer_input = args[:layer_length]
original_inputs = args[layer_length:-1]
device_ids = self.device_ids
if device_ids is None:
device_ids = getattr(self.forward_func, "device_ids", None)
all_layer_inputs = {}
if device_ids is not None:
scattered_layer_input = scatter(layer_input,
target_gpus=device_ids)
for device_tensors in scattered_layer_input:
all_layer_inputs[device_tensors[0].device] = device_tensors
else:
all_layer_inputs[layer_input[0].device] = layer_input
def forward_hook(module, inp, out=None):
device = _extract_device(module, inp, out)
is_layer_tuple = (isinstance(out, tuple) if out is not None
else isinstance(inp, tuple))
if device not in all_layer_inputs:
raise AssertionError(
"Layer input not placed on appropriate "
"device. If using a DataParallel model, either provide the "
"DataParallel model as forward_func or provide device ids"
" to the constructor.")
if not is_layer_tuple:
return all_layer_inputs[device][0]
return all_layer_inputs[device]
hook = None
try:
if attribute_to_layer_input:
hook = self.layer.register_forward_pre_hook(forward_hook)
else:
hook = self.layer.register_forward_hook(forward_hook)
eval = _run_forward(self.forward_func,
original_inputs,
target=target)
finally:
if hook is not None:
hook.remove()
return eval
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 _forward_layer_distributed_eval(
forward_fn: Callable,
inputs: Union[Tensor, Tuple[Tensor, ...]],
layer: ModuleOrModuleList,
target_ind: TargetType = None,
additional_forward_args: Any = None,
attribute_to_layer_input: bool = False,
forward_hook_with_return: bool = False,
require_layer_grads: bool = False,
) -> Union[
Tuple[Dict[Module, Dict[device, Tuple[Tensor, ...]]], Tensor],
Dict[Module, Dict[device, Tuple[Tensor, ...]]],
]:
r"""
A helper function that allows to set a hook on model's `layer`, run the forward
pass and returns intermediate layer results, stored in a dictionary,
and optionally also the output of the forward function. The keys in the
dictionary are the device ids and the values are corresponding intermediate layer
results, either the inputs or the outputs of the layer depending on whether we set
`attribute_to_layer_input` to True or False.
This is especially useful when we execute forward pass in a distributed setting,
using `DataParallel`s for example.
"""
saved_layer: Dict[Module, Dict[device, Tuple[Tensor, ...]]] = defaultdict(dict)
lock = threading.Lock()
all_layers: List[Module] = [layer] if isinstance(layer, Module) else layer
# Set a forward hook on specified module and run forward pass to
# get layer output tensor(s).
# For DataParallel models, each partition adds entry to dictionary
# with key as device and value as corresponding Tensor.
def hook_wrapper(original_module):
def forward_hook(module, inp, out=None):
eval_tsrs = inp if attribute_to_layer_input else out
is_eval_tuple = isinstance(eval_tsrs, tuple)
if not is_eval_tuple:
eval_tsrs = (eval_tsrs,)
if require_layer_grads:
apply_gradient_requirements(eval_tsrs, warn=False)
with lock:
nonlocal saved_layer
# Note that cloning behaviour of `eval_tsr` is different
# when `forward_hook_with_return` is set to True. This is because
# otherwise `backward()` on the last output layer won't execute.
if forward_hook_with_return:
saved_layer[original_module][eval_tsrs[0].device] = eval_tsrs
eval_tsrs_to_return = tuple(
eval_tsr.clone() for eval_tsr in eval_tsrs
)
if not is_eval_tuple:
eval_tsrs_to_return = eval_tsrs_to_return[0]
return eval_tsrs_to_return
else:
saved_layer[original_module][eval_tsrs[0].device] = tuple(
eval_tsr.clone() for eval_tsr in eval_tsrs
)
return forward_hook
all_hooks = []
try:
for single_layer in all_layers:
if attribute_to_layer_input:
all_hooks.append(
single_layer.register_forward_pre_hook(hook_wrapper(single_layer))
)
else:
all_hooks.append(
single_layer.register_forward_hook(hook_wrapper(single_layer))
)
output = _run_forward(
forward_fn,
inputs,
target=target_ind,
additional_forward_args=additional_forward_args,
)
finally:
for hook in all_hooks:
hook.remove()
if len(saved_layer) == 0:
raise AssertionError("Forward hook did not obtain any outputs for given layer")
if forward_hook_with_return:
return saved_layer, output
return saved_layer
def _calculate_vis_output(self,
inputs,
additional_forward_args,
label,
target=None) -> Optional[VisualizationOutput]:
net = self.models[0] # TODO process multiple models
# initialize baselines
baseline_transforms_len = len(self.features[0].baseline_transforms)
baselines: List[List[Optional[Tensor]]] = [
[None] * len(self.features) for _ in range(baseline_transforms_len)
]
transformed_inputs = list(inputs)
# transformed_inputs = list([i.clone() for i in inputs])
for feature_i, feature in enumerate(self.features):
transformed_inputs[feature_i] = self._transform(
feature.input_transforms, transformed_inputs[feature_i], True)
assert baseline_transforms_len == len(
feature.baseline_transforms
), "Must have same number of baselines across all features"
for baseline_i, baseline_transform in enumerate(
feature.baseline_transforms):
baselines[baseline_i][feature_i] = self._transform(
[baseline_transform], transformed_inputs[feature_i], True)
baselines = cast(List[List[Tensor]], baselines)
outputs = _run_forward(
net,
tuple(transformed_inputs),
additional_forward_args=additional_forward_args,
)
if self.score_func is not None:
outputs = self.score_func(outputs)
if outputs.nelement() == 1:
scores = outputs
predicted = scores.round().to(torch.int)
else:
scores, predicted = outputs.topk(min(4, outputs.shape[-1]))
scores = scores.cpu().squeeze(0)
predicted = predicted.cpu().squeeze(0)
actual_label_output = None
if label is not None and len(label) > 0:
actual_label_output = OutputScore(score=100,
index=label[0],
label=self.classes[label[0]])
predicted_scores = self._get_labels_from_scores(scores, predicted)
# Filter based on UI configuration
if actual_label_output is None or not self._should_keep_prediction(
predicted_scores, actual_label_output):
return None
baselines_group = [tuple(b) for b in baselines]
if target is None:
target = predicted_scores[0].index if len(
predicted_scores) > 0 else None
# attributions are given per input*
# inputs given to the model are described via `self.features`
#
# *an input contains multiple features that represent it
# e.g. all the pixels that describe an image is an input
attrs_per_input_feature = self._calculate_attribution(
net,
baselines_group,
tuple(transformed_inputs),
additional_forward_args,
target,
)
net_contrib = self._calculate_net_contrib(attrs_per_input_feature)
# the features per input given
features_per_input = [
feature.visualize(attr, data, contrib)
for feature, attr, data, contrib in zip(
self.features, attrs_per_input_feature, inputs, net_contrib)
]
return VisualizationOutput(
feature_outputs=features_per_input,
actual=actual_label_output,
predicted=predicted_scores,
active_index=target
if target is not None else actual_label_output.index,
)
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,
**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
**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_input(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():
# 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
)
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._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,
)
# (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,))
if self.use_weights:
weights[i] += current_mask.float().sum(dim=0)
total_attrib[i] += (eval_diff * current_mask.to(attrib_type)).sum(
dim=0
)
# 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 compute_convergence_delta(
self,
attributions: Union[Tensor, Tuple[Tensor, ...]],
start_point: Union[None, int, float, Tensor,
Tuple[Union[int, float, Tensor], ...]],
end_point: Union[Tensor, Tuple[Tensor, ...]],
target: TargetType = None,
additional_forward_args: Any = None,
) -> Tensor:
r"""
Here we provide a specific implementation for `compute_convergence_delta`
which is based on a common property among gradient-based attribution algorithms.
In the literature sometimes it is also called completeness axiom. Completeness
axiom states that the sum of the attribution must be equal to the differences of
NN Models's function at its end and start points. In other words:
sum(attributions) - (F(end_point) - F(start_point)) is close to zero.
Returned delta of this method is defined as above stated difference.
This implementation assumes that both the `start_point` and `end_point` have
the same shape and dimensionality. It also assumes that the target must have
the same number of examples as the `start_point` and the `end_point` in case
it is provided in form of a list or a non-singleton tensor.
Args:
attributions (tensor or tuple of tensors): Precomputed attribution
scores. The user can compute those using any attribution
algorithm. It is assumed the the shape and the
dimensionality of attributions must match the shape and
the dimensionality of `start_point` and `end_point`.
It also assumes that the attribution tensor's
dimension 0 corresponds to the number of
examples, and if multiple input tensors are provided,
the examples must be aligned appropriately.
start_point (tensor or tuple of tensors, optional): `start_point`
is passed as an input to model's forward function. It
is the starting point of attributions' approximation.
It is assumed that both `start_point` and `end_point`
have the same shape and dimensionality.
end_point (tensor or tuple of tensors): `end_point`
is passed as an input to model's forward function. It
is the end point of attributions' approximation.
It is assumed that both `start_point` and `end_point`
have the same shape and dimensionality.
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.
`additional_forward_args` is used both for `start_point`
and `end_point` when computing the forward pass.
Default: None
Returns:
*tensor* of **deltas**:
- **deltas** (*tensor*):
This implementation returns convergence delta per
sample. Deriving sub-classes may do any type of aggregation
of those values, if necessary.
"""
end_point, start_point = _format_input_baseline(end_point, start_point)
additional_forward_args = _format_additional_forward_args(
additional_forward_args)
# tensorizing start_point in case it is a scalar or one example baseline
# If the batch size is large we could potentially also tensorize only one
# sample and expand the output to the rest of the elements in the batch
start_point = _tensorize_baseline(end_point, start_point)
attributions = _format_tensor_into_tuples(attributions)
# verify that the attributions and end_point match on 1st dimension
for attribution, end_point_tnsr in zip(attributions, end_point):
assert end_point_tnsr.shape[0] == attribution.shape[0], (
"Attributions tensor and the end_point must match on the first"
" dimension but found attribution: {} and end_point: {}".
format(attribution.shape[0], end_point_tnsr.shape[0]))
num_samples = end_point[0].shape[0]
_validate_input(end_point, start_point)
_validate_target(num_samples, target)
with torch.no_grad():
start_out_sum = _sum_rows(
_run_forward(self.forward_func, start_point, target,
additional_forward_args))
end_out_sum = _sum_rows(
_run_forward(self.forward_func, end_point, target,
additional_forward_args))
row_sums = [_sum_rows(attribution) for attribution in attributions]
attr_sum = torch.stack(
[cast(Tensor, sum(row_sum)) for row_sum in zip(*row_sums)])
_delta = attr_sum - (end_out_sum - start_out_sum)
return _delta
def attribute(
self,
inputs: TensorOrTupleOfTensorsGeneric,
baselines: BaselineType = None,
target: TargetType = None,
additional_forward_args: Any = None,
feature_mask: Union[None, TensorOrTupleOfTensorsGeneric] = None,
n_samples: int = 25,
perturbations_per_eval: int = 1,
show_progress: bool = False,
) -> TensorOrTupleOfTensorsGeneric:
r"""
NOTE: The feature_mask argument differs from other perturbation based
methods, since feature indices can overlap across tensors. See the
description of the feature_mask argument below for more details.
Args:
inputs (tensor or tuple of tensors): Input for which Shapley value
sampling 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 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 difference is computed (for classification cases,
this is usually the target class).
If the network returns a scalar value per example,
no target index is necessary.
For general 2D outputs, targets can be either:
- a single integer or a tensor containing a single
integer, which is applied to all input examples
- a list of integers or a 1D tensor, with length matching
the number of examples in inputs (dim 0). Each integer
is applied as the target for the corresponding example.
For outputs with > 2 dimensions, targets can be either:
- A single tuple, which contains #output_dims - 1
elements. This target index is applied to all examples.
- A list of tuples with length equal to the number of
examples in inputs (dim 0), and each tuple containing
#output_dims - 1 elements. Each tuple is applied as the
target for the corresponding example.
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 added 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. Values across
all tensors should be integers in the range 0 to
num_features - 1, and indices corresponding to the same
feature should have the same value.
Note that features are grouped across tensors
(unlike feature ablation and occlusion), so
if the same index is used in different tensors, those
features are still grouped and added simultaneously.
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
Default: None
n_samples (int, optional): The number of feature permutations
tested.
Default: `25` if `n_samples` is not provided.
perturbations_per_eval (int, optional): Allows multiple ablations
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 returns a single scalar per batch,
perturbations_per_eval must be set to 1.
Default: 1
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
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 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 ShapleyValueSampling interpreter
>>> svs = ShapleyValueSampling(net)
>>> # Computes attribution, taking random orderings
>>> # of the 16 features and computing the output change when adding
>>> # each feature. We average over 200 trials (random permutations).
>>> attr = svs.attribute(input, target=1, n_samples=200)
>>> # Alternatively, we may want to add features in groups, e.g.
>>> # grouping each 2x2 square of the inputs and adding 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 added
>>> # together, 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 = svs.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
)
feature_mask = (
_format_tensor_into_tuples(feature_mask)
if feature_mask is not None
else None
)
assert (
isinstance(perturbations_per_eval, int) and perturbations_per_eval >= 1
), "Ablations per evaluation must be at least 1."
with torch.no_grad():
baselines = _tensorize_baseline(inputs, baselines)
num_examples = inputs[0].shape[0]
if feature_mask is None:
feature_mask, total_features = _construct_default_feature_mask(inputs)
else:
total_features = int(
max(torch.max(single_mask).item() for single_mask in feature_mask)
+ 1
)
if show_progress:
attr_progress = progress(
desc=f"{self.get_name()} attribution",
total=self._get_n_evaluations(
total_features, n_samples, perturbations_per_eval
)
+ 1, # add 1 for the initial eval
)
attr_progress.update(0)
initial_eval = _run_forward(
self.forward_func, baselines, target, additional_forward_args
)
if show_progress:
attr_progress.update()
agg_output_mode = _find_output_mode_and_verify(
initial_eval, num_examples, perturbations_per_eval, feature_mask
)
# Initialize attribution totals and counts
total_attrib = [
torch.zeros_like(
input[0:1] if agg_output_mode else input, dtype=torch.float
)
for input in inputs
]
iter_count = 0
# Iterate for number of samples, generate a permutation of the features
# and evalute the incremental increase for each feature.
for feature_permutation in self.permutation_generator(
total_features, n_samples
):
iter_count += 1
prev_results = initial_eval
for (
current_inputs,
current_add_args,
current_target,
current_masks,
) in self._perturbation_generator(
inputs,
additional_forward_args,
target,
baselines,
feature_mask,
feature_permutation,
perturbations_per_eval,
):
if sum(torch.sum(mask).item() for mask in current_masks) == 0:
warnings.warn(
"Feature mask is missing some integers between 0 and "
"num_features, for optimal performance, make sure each"
" consecutive integer corresponds to a feature."
)
# 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()
if agg_output_mode:
eval_diff = modified_eval - prev_results
prev_results = modified_eval
else:
all_eval = torch.cat((prev_results, modified_eval), dim=0)
eval_diff = all_eval[num_examples:] - all_eval[:-num_examples]
prev_results = all_eval[-num_examples:]
for j in range(len(total_attrib)):
current_eval_diff = eval_diff
if not agg_output_mode:
# current_eval_diff dimensions:
# (#features in batch, #num_examples, 1,.. 1)
# (contains 1 more dimension than inputs). This adds extra
# dimensions of 1 to make the tensor broadcastable with the
# inputs tensor.
current_eval_diff = current_eval_diff.reshape(
(-1, num_examples) + (len(inputs[j].shape) - 1) * (1,)
)
total_attrib[j] += (
current_eval_diff * current_masks[j].float()
).sum(dim=0)
if show_progress:
attr_progress.close()
# Divide total attributions by number of random permutations and return
# formatted attributions.
attrib = tuple(
tensor_attrib_total / iter_count for tensor_attrib_total in total_attrib
)
formatted_attr = _format_output(is_inputs_tuple, attrib)
return formatted_attr
