Lucid3.0
fn

register_module_full_backward_hook

→RemovableHandle
register_module_full_backward_hook(hook: Callable[..., object])
source

Register a global post-hook fired after a module's backward pass.

Invoked once Lucid's autograd engine has finished accumulating gradients for a Module. The hook receives the module along with the input-gradients (gradients flowing back into the module's inputs) and the output-gradients (gradients that came from above). Mutating these in place propagates to subsequent backward computation; returning a new tuple of input-gradients replaces them.

Parameters

hookCallable
Signature: hook(module, grad_input, grad_output) -> None | tuple.

Returns

RemovableHandle

Handle for later deregistration.

Notes

Use the "full" variant whenever a Module has multiple inputs and you need access to every per-input gradient. Pairs naturally with register_module_full_backward_pre_hook for modify-then-observe workflows.

Examples

>>> from lucid.nn.hooks import register_module_full_backward_hook
>>> def log_grad_norm(mod, grad_input, grad_output):
...     for i, g in enumerate(grad_input):
...         if g is not None:
...             print(f'{type(mod).__name__}.in[{i}].grad_norm = {g.norm().item()}')
>>> handle = register_module_full_backward_hook(log_grad_norm)