lucid.autograd

The lucid.autograd package provides explicit autograd utilities that mirror common PyTorch workflows while keeping gradient calls simple and transparent. Use this package when you want gradients returned as values or when you need to trigger the backward engine directly without calling Tensor methods.

Overview

The package offers two primary entry points:

  • grad: Compute gradients of outputs with respect to inputs and return them without leaving gradients on the input tensors.

  • backward: Run the autograd engine explicitly, mirroring the behavior of Tensor.backward while centralizing the backward logic.

Key Features

Functional Gradient Extraction

lucid.autograd.grad is useful for functional-style training loops, custom loss aggregation, or when you want to avoid mutating .grad on inputs.

Example

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

Explicit Engine Control

lucid.autograd.backward provides direct access to the backward engine. This is helpful for higher-level utilities that want to orchestrate backward passes without tying logic to the Tensor class.

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.])

Usage Notes

  • grad_outputs can be used to seed gradients for vector-Jacobian products.

  • By default, intermediate gradients are cleared unless retain_grad is set.

  • The autograd utilities track first-order gradients only.

See Also

  • lucid.autograd.grad

  • lucid.autograd.backward