lucid.autograd.backward¶

lucid.autograd.backward(tensor: _TensorLike, retain_grad: bool = False, retain_graph: bool = False) → None¶

lucid.autograd.backward runs the internal autograd engine, mirroring the behavior of Tensor.backward while centralizing the backward logic. It is useful when you need to explicitly invoke the engine without calling the Tensor method.

Function Signature¶

def backward(
    tensor: Tensor,
    retain_grad: bool = False,
    retain_graph: bool = False,
) -> None

Parameters¶

tensor: The output Tensor to backpropagate from.
retain_grad: Whether to keep non-leaf gradients.
retain_graph: Whether to keep the graph for further backward passes.

Behavior Notes¶

Seeds the output gradient with ones if tensor.grad is not already set.
Clears intermediate nodes and operation results unless retain_graph is set.
Mirrors the same behavior as Tensor.backward internally.

Example¶

>>> import lucid
>>> x = lucid.tensor([1.0, 2.0, 3.0], requires_grad=True)
>>> y = lucid.sum(x * x)
>>> lucid.autograd.backward(y)
>>> x.grad
array([2., 4., 6.])

Manual Seed Gradient¶

>>> import numpy as np
>>> x = lucid.tensor([1.0, 2.0, 3.0], requires_grad=True)
>>> y = lucid.sum(x * x)
>>> y.grad = np.array([10.0])
>>> lucid.autograd.backward(y)
>>> x.grad
array([20., 40., 60.])