Core: Multiple Inputs and Keyword Arguments

- use additions to forward hooks in torch 2.0.0 to pass kwargs to
  pass keyword arguments
- handle multiple inputs and outputs in core.Hook and core.BasicHook, by
  passing all required grad_outputs and inputs to the backward
  implementation

TODO:

- attribution scores are currently wrong in BasicHook, likely an issue
  with the gradient inside BasicHook? Might be some cross-terms
  interacting that should not interact

- finish draft and test implementation
- add tests
- add documentation

- This stands in conflict with #168, but promises a better
  implementation by handling inputs and outpus as common to a single
  function, rather than individually as proposed in #168

Author: chr5tphr

setup.py

@@ -54,7 +54,7 @@ def replace(mobj):
         'click',
         'numpy',
         'Pillow',
-        'torch>=1.7.0',
+        'torch>=2.0.0',
         'torchvision',
     ],
     setup_requires=[

src/zennit/core.py

@@ -19,6 +19,8 @@
 import functools
 import weakref
 from contextlib import contextmanager
+from itertools import compress, repeat, islice, chain
+from inspect import signature
 
 import torch
 
@@ -234,6 +236,43 @@ def modifier_wrapper(input, name):
     return zero_params_wrapper
 
 
+def uncompress(data, selector, compressed):
+    '''Generator which, given a compressed iterable produced by :py:obj:`itertools.compress` and (some iterable similar
+    to) the original data and selector used for :py:obj:`~itertools.compress`, yields values from `compressed` or
+    `data` depending on `selector`. `True` values in `selector` skip `data` one ahead and yield a value from
+    `compressed`, while `False` values yield one value from `data`.
+
+    Parameters
+    ----------
+    data : iterable
+        The iterable (similar to the) original data. `False` values in the `selector` will be filled with values from
+        this iterator, while `True` values will cause this iterable to be skipped.
+    selector : iterable of bool
+        The original selector used to produce `compressed`. Chooses whether elements from `data` or from `compressed`
+        will be yielded.
+    compressed : iterable
+        The results of :py:obj:`itertools.compress`. Will be yielded for each `True` element in `selector`.
+
+    Yields
+    ------
+    object
+        An element of `data` if the associated element of `selector` is `False`, otherwise an element of `compressed`
+        while skipping `data` one ahead.
+
+    '''
+    its = iter(selector)
+    itc = iter(compressed)
+    itd = iter(data)
+    try:
+        if next(its):
+            next(itd)
+            yield next(itc)
+        else:
+            yield next(itd)
+    except StopIteration:
+        return
+
+
 class ParamMod:
     '''Class to produce a context manager to temporarily modify parameter attributes (all by default) of a module.
 
@@ -360,6 +399,7 @@ class Identity(torch.autograd.Function):
     @staticmethod
     def forward(ctx, *inputs):
         '''Forward identity.'''
+        ctx.mark_non_differentiable(*[elem for elem in inputs if not elem.requires_grad])
         return inputs
 
     @staticmethod
@@ -375,62 +415,94 @@ def __init__(self):
         self.active = True
         self.tensor_handles = RemovableHandleList()
 
-    def pre_forward(self, module, input):
+    @staticmethod
+    def _inject_grad_fn(args):
+        tensor_mask = tuple(isinstance(elem, torch.Tensor) for elem in args)
+        tensors = tuple(compress(args, tensor_mask))
+        # tensors = [(n, elem) for elem in enumerate(args) if isinstance(elem, torch.Tensor)]
+
+        # only if gradient required
+        if not any(tensor.requires_grad for tensor in tensors):
+            return None, args, tensor_mask
+
+        # add identity to ensure .grad_fn exists and all tensors share the same .grad_fn
+        post_tensors = Identity.apply(*tensors)
+        grad_fn = next((tensor.grad_fn for tensor in post_tensors if tensor.grad_fn is not None), None)
+        if grad_fn is None:
+            raise RuntimeError('Backward hook could not be registered!')
+
+        # work-around to support in-place operations
+        post_tensors = tuple(elem.clone() for elem in post_tensors)
+        post_args = tuple(uncompress(args, tensor_mask, post_tensors))
+        return grad_fn, post_args, tensor_mask
+
+    def pre_forward(self, module, args, kwargs):
         '''Apply an Identity to the input before the module to register a backward hook.'''
         hook_ref = weakref.ref(self)
 
+        grad_fn, post_args, input_tensor_mask = self._inject_grad_fn(args)
+        if grad_fn is None:
+            return
+
         @functools.wraps(self.backward)
         def wrapper(grad_input, grad_output):
             hook = hook_ref()
             if hook is not None and hook.active:
-                return hook.backward(module, grad_input, hook.stored_tensors['grad_output'])
+                return hook.backward(
+                    module,
+                    list(uncompress(
+                        repeat(None),
+                        input_tensor_mask,
+                        grad_input,
+                    )),
+                    hook.stored_tensors['grad_output'],
+                )
             return None
 
-        if not isinstance(input, tuple):
-            input = (input,)
+        # register the input tensor gradient hook
+        self.tensor_handles.append(grad_fn.register_hook(wrapper))
 
-        # only if gradient required
-        if input[0].requires_grad:
-            # add identity to ensure .grad_fn exists
-            post_input = Identity.apply(*input)
-            # register the input tensor gradient hook
-            self.tensor_handles.append(
-                post_input[0].grad_fn.register_hook(wrapper)
-            )
-            # work around to support in-place operations
-            post_input = tuple(elem.clone() for elem in post_input)
-        else:
-            # no gradient required
-            post_input = input
-        return post_input[0] if len(post_input) == 1 else post_input
+        return post_args, kwargs
 
-    def post_forward(self, module, input, output):
+    def post_forward(self, module, args, kwargs, output):
         '''Register a backward-hook to the resulting tensor right after the forward.'''
         hook_ref = weakref.ref(self)
 
+        single = not isinstance(output, tuple)
+        if single:
+            output = (output,)
+
+        grad_fn, post_output, output_tensor_mask = self._inject_grad_fn(output)
+        if grad_fn is None:
+            return
+
         @functools.wraps(self.pre_backward)
         def wrapper(grad_input, grad_output):
             hook = hook_ref()
             if hook is not None and hook.active:
-                return hook.pre_backward(module, grad_input, grad_output)
+                return hook.pre_backward(
+                    module,
+                    grad_input,
+                    tuple(uncompress(
+                        repeat(None),
+                        output_tensor_mask,
+                        grad_output
+                    ))
+                )
             return None
 
-        if not isinstance(output, tuple):
-            output = (output,)
+        # register the output tensor gradient hook
+        self.tensor_handles.append(grad_fn.register_hook(wrapper))
 
-        # only if gradient required
-        if output[0].grad_fn is not None:
-            # register the output tensor gradient hook
-            self.tensor_handles.append(
-                output[0].grad_fn.register_hook(wrapper)
-            )
-        return output[0] if len(output) == 1 else output
+        if single:
+            return post_output[0]
+        return post_output
 
     def pre_backward(self, module, grad_input, grad_output):
         '''Store the grad_output for the backward hook'''
         self.stored_tensors['grad_output'] = grad_output
 
-    def forward(self, module, input, output):
+    def forward(self, module, args, kwargs, output):
         '''Hook applied during forward-pass'''
 
     def backward(self, module, grad_input, grad_output):
@@ -449,11 +521,14 @@ def remove(self):
 
     def register(self, module):
         '''Register this instance by registering all hooks to the supplied module.'''
+        # assume with_kwargs if forward has not 3 parameters and 3rd is not called 'output'
+        forward_params = signature(self.forward).parameters
+        with_kwargs = len(forward_params) != 3 and list(forward_params)[2] != 'output'
         return RemovableHandleList([
             RemovableHandle(self),
-            module.register_forward_pre_hook(self.pre_forward),
-            module.register_forward_hook(self.post_forward),
-            module.register_forward_hook(self.forward),
+            module.register_forward_pre_hook(self.pre_forward, with_kwargs=True),
+            module.register_forward_hook(self.post_forward, with_kwargs=True),
+            module.register_forward_hook(self.forward, with_kwargs=with_kwargs),
         ])
 
 
@@ -517,31 +592,61 @@ def __init__(
         self.gradient_mapper = gradient_mapper
         self.reducer = reducer
 
-    def forward(self, module, input, output):
+    def forward(self, module, args, kwargs, output):
         '''Forward hook to save module in-/outputs.'''
-        self.stored_tensors['input'] = input
+        self.stored_tensors['input'] = args
+        self.stored_tensors['kwargs'] = kwargs
 
     def backward(self, module, grad_input, grad_output):
         '''Backward hook to compute LRP based on the class attributes.'''
-        original_input = self.stored_tensors['input'][0].clone()
+        input_mask = [elem is not None for elem in self.stored_tensors['input']]
+        output_mask = [elem is not None for elem in grad_output]
+        cgrad_output = tuple(compress(grad_output, output_mask))
+
+        original_inputs = [tensor.clone() for tensor in self.stored_tensors['input']]
+        kwargs = self.stored_tensors['kwargs']
         inputs = []
         outputs = []
         for in_mod, param_mod, out_mod in zip(self.input_modifiers, self.param_modifiers, self.output_modifiers):
-            input = in_mod(original_input).requires_grad_()
+            mod_args = (in_mod(tensor).requires_grad_() for tensor in compress(original_inputs, input_mask))
+            args = tuple(uncompress(original_inputs, input_mask, mod_args))
             with ParamMod.ensure(param_mod)(module) as modified, torch.autograd.enable_grad():
-                output = modified.forward(input)
-                output = out_mod(output)
-            inputs.append(input)
+                output = modified.forward(*args, **kwargs)
+                if not isinstance(output, tuple):
+                    output = (output,)
+                output = tuple(out_mod(tensor) for tensor in compress(output, output_mask))
+            inputs.append(compress(args, input_mask))
             outputs.append(output)
-        grad_outputs = self.gradient_mapper(grad_output[0], outputs)
-        gradients = torch.autograd.grad(
-            outputs,
-            inputs,
+
+        inputs = list(zip(*inputs))
+        outputs = list(zip(*outputs))
+        input_struct = [len(elem) for elem in inputs]
+        output_struct = [len(elem) for elem in outputs]
+
+        grad_outputs = tuple(
+            self.gradient_mapper(gradout, outs)
+            for gradout, outs in zip(cgrad_output, outputs)
+        )
+        inputs_flat = tuple(chain.from_iterable(inputs))
+        outputs_flat = tuple(chain.from_iterable(outputs))
+        if not all(isinstance(elem, torch.Tensor) for elem in grad_outputs):
+            # if there is only a single output modifier, grad_outputs may contain tensors
+            grad_outputs = tuple(chain.from_iterable(grad_outputs))
+
+        gradients_flat = torch.autograd.grad(
+            outputs_flat,
+            inputs_flat,
             grad_outputs=grad_outputs,
-            create_graph=grad_output[0].requires_grad
+            create_graph=any(tensor.requires_grad for tensor in cgrad_output)
         )
-        relevance = self.reducer(inputs, gradients)
-        return tuple(relevance if original.shape == relevance.shape else None for original in grad_input)
+
+        # input_it = iter(inputs)
+        # inputs_re = [tuple(islice(input_it, size)) for size in input_struct]
+        gradient_it = iter(gradients_flat)
+        gradients = [tuple(islice(gradient_it, size)) for size in input_struct]
+
+        relevances = (self.reducer(inp, grad) for inp, grad in zip(inputs, gradients))
+        return tuple(uncompress(repeat(None), input_mask, relevances))
 
     def copy(self):
         '''Return a copy of this hook.