tdhook.attribution.lrp_helpers.rules

LRP rules.

This code is adapted from the Zennit library (LGPL-3.0) and the LXT library (Clear BSD) Original sources: - chr5tphr/zennit - rachtibat/LRP-eXplains-Transformers

Classes

ParamModifier
RemovableRuleHandle
AbstractFunctionMeta	Metaclass for defining Abstract Base Classes (ABCs).
Rule	Base class to create custom autograd.Function.
EpsilonRule	Base class to create custom autograd.Function.
UniformEpsilonRule	Base class to create custom autograd.Function.
PassRule	Base class to create custom autograd.Function.
IgnoreRule	Base class to create custom autograd.Function.
WSquareRule	Base class to create custom autograd.Function.
FlatRule	Base class to create custom autograd.Function.
UniformRule	Base class to create custom autograd.Function.
StopRule	Base class to create custom autograd.Function.
AlphaBetaRule	Base class to create custom autograd.Function.
SoftmaxEpsilonRule	Base class to create custom autograd.Function.
LayerNormRule	Base class to create custom autograd.Function.
PseudoIdentityRule	Base class to create custom autograd.Function.
AHQKVRule	Base class to create custom autograd.Function.
BaseRuleMapper
EpsilonPlus

Functions

stabilize(tensor[, epsilon])
raise_for_unconserved_rel_factory([atol, rtol])

Module Contents

tdhook.attribution.lrp_helpers.rules.stabilize(tensor, epsilon=1e-06)

class tdhook.attribution.lrp_helpers.rules.ParamModifier(modify_fn=None, select_fn=None)

Parameters:

• modify_fn (Optional[Callable[[str, torch.nn.Parameter], torch.Tensor]])

• select_fn (Optional[Callable[[str, torch.nn.Parameter], bool]])

_modify_fn

_select_fn

state_dicts(module)

Parameters:

module (torch.nn.Module)

__call__(module)

Parameters:

module (torch.nn.Module)

static from_modifiers(modifiers)

Parameters:

modifiers (List[ParamModifier])

static select_all(name, param)

Parameters:

• name (str)

• param (torch.nn.Parameter)

class tdhook.attribution.lrp_helpers.rules.RemovableRuleHandle(rule, module)

Parameters:

• rule (Rule)

• module (torch.nn.Module)

_rule

_module_ref

__enter__()

__exit__(exc_type, exc_value, traceback)

remove()

class tdhook.attribution.lrp_helpers.rules.AbstractFunctionMeta(name, bases, attrs)

Bases: abc.ABCMeta, torch.autograd.function.FunctionMeta

Metaclass for defining Abstract Base Classes (ABCs).

Use this metaclass to create an ABC. An ABC can be subclassed directly, and then acts as a mix-in class. You can also register unrelated concrete classes (even built-in classes) and unrelated ABCs as ‘virtual subclasses’ – these and their descendants will be considered subclasses of the registering ABC by the built-in issubclass() function, but the registering ABC won’t show up in their MRO (Method Resolution Order) nor will method implementations defined by the registering ABC be callable (not even via super()).

class tdhook.attribution.lrp_helpers.rules.Rule

Bases: torch.autograd.function.Function

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

_apply_kwargs

register(module)

Parameters:

module (torch.nn.Module)

unregister(module)

Parameters:

module (torch.nn.Module)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Abstractmethod:

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevance)

Abstractmethod:

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.EpsilonRule(epsilon=1e-06)

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

property epsilon

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevance)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.UniformEpsilonRule(epsilon=1e-06)

Bases: EpsilonRule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static backward(ctx, *out_relevance)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.PassRule

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevance)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.IgnoreRule

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

register(module)

Parameters:

module (torch.nn.Module)

unregister(module)

Parameters:

module (torch.nn.Module)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevance)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.WSquareRule(stabilizer=1e-06)

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevance)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.FlatRule(stabilizer=1e-06)

Bases: WSquareRule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

class tdhook.attribution.lrp_helpers.rules.UniformRule

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevances)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.StopRule

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevances)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.AlphaBetaRule(alpha=2.0, beta=1.0, stabilizer=1e-06)

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevance)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.SoftmaxEpsilonRule(epsilon=1e-06)

Bases: EpsilonRule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static backward(ctx, *out_relevances)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.LayerNormRule

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevances)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.PseudoIdentityRule(stabilizer=1e-06)

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevances)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.AHQKVRule

Bases: Rule

Base class to create custom autograd.Function.

To create a custom autograd.Function, subclass this class and implement the forward() and backward() static methods. Then, to use your custom op in the forward pass, call the class method apply. Do not call forward() directly.

To ensure correctness and best performance, make sure you are calling the correct methods on ctx and validating your backward function using torch.autograd.gradcheck().

See extending-autograd for more details on how to use this class.

Examples:

>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_AUTOGRAD)
>>> class Exp(Function):
>>>     @staticmethod
>>>     def forward(ctx, i):
>>>         result = i.exp()
>>>         ctx.save_for_backward(result)
>>>         return result
>>>
>>>     @staticmethod
>>>     def backward(ctx, grad_output):
>>>         result, = ctx.saved_tensors
>>>         return grad_output * result
>>>
>>> # Use it by calling the apply method:
>>> # xdoctest: +SKIP
>>> output = Exp.apply(input)

static forward(ctx, apply_kwargs, module, model_kwargs, *inputs)

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, *out_relevances)

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class tdhook.attribution.lrp_helpers.rules.BaseRuleMapper(stabilizer=1e-06, rule_mapper=None)

Parameters:

rule_mapper (Optional[Callable[[str, torch.nn.Module], Rule | None]])

_stabilizer = 1e-06

_rule_mapper

_rules

_call(name, module)

Parameters:

• name (str)

• module (torch.nn.Module)

Return type:

Rule | None

__call__(name, module)

Parameters:

• name (str)

• module (torch.nn.Module)

Return type:

Rule | None

class tdhook.attribution.lrp_helpers.rules.EpsilonPlus(epsilon=1e-06, stabilizer=1e-06, rule_mapper=None)

Bases: BaseRuleMapper

Parameters:

rule_mapper (Optional[Callable[[str, torch.nn.Module], Rule | None]])

_call(name, module)

Parameters:

• name (str)

• module (torch.nn.Module)

Return type:

Rule | None

tdhook.attribution.lrp_helpers.rules.raise_for_unconserved_rel_factory(atol=1e-06, rtol=1e-06)

Parameters:

• atol (float)

• rtol (float)