Lucid3.0
module

autograd

18 members
lucid.autograd

Classes

classno_grad3 methods

Context manager / decorator that disables gradient tracking.

classenable_grad3 methods

Context manager / decorator that (re-)enables gradient tracking.

classFunction3 methods

Base class for custom differentiable operations.

classFunctionCtx4 methods

Per-call context shared between `Function.forward` and `Function.backward`.

classdetect_anomaly4 methods

Context manager / decorator that enables autograd anomaly detection.

classRemovableHandle4 methods

Handle returned by `lucid.Tensor.register_hook`.

Functions

fnset_grad_enabled→ None

Globally set the autograd gradient-tracking flag.

fnis_grad_enabled→ bool

Return whether autograd gradient tracking is currently enabled.

fninference_mode→ Iterator[None]

Context manager for inference-time autograd suppression.

fnbackward→ None

Compute gradients of `tensors` w.r.t. the leaf variables in their graph.

fngrad→ tuple[Tensor or None, ...]

Compute gradients of outputs w.r.t. inputs, returning them as a tuple.

fngradgradcheck→ bool

Verify second-order gradients via finite differences.

fnis_anomaly_enabled→ bool

Return whether autograd anomaly detection is currently enabled.

fnset_detect_anomaly→ None

Programmatic global toggle for autograd anomaly detection.

fnjacobian→ Tensor or tuple of Tensor

Compute the Jacobian matrix of `func` with respect to each input.

fnhessian→ Tensor or tuple of tuple of Tensor

Compute the Hessian matrix of a scalar-valued `func`.

fnvjp→ tuple of (Tensor, tuple of (Tensor or None))

Vector-Jacobian product $v^\top J$ (reverse-mode AD).

fnjvp→ tuple of (Tensor or tuple of Tensor, Tensor or tuple of Tensor)

Jacobian-vector product $J v$ (forward-mode directional derivative).