Autograd #

Autograd widgets (tapes + gradients).

This module provides infoview panels for TorchLean’s eager-mode autograd tape:

#tape_view t renders the tape nodes, showing parents, stored forward values, and a DOT snippet.
#tape_grads_view t, outId runs reverse-mode from a scalar output (outId) and shows which node ids received gradients, plus small previews.

Main definitions #

tapeHtml: inspect a tape's nodes and edge structure.
tapeGradsHtml: run scalar reverse-mode and show gradient coverage/results.
tapeTraceHtml: accumulated cotangents and recomputed per-parent VJP contributions.
#tape_view, #tape_grads_view, #tape_trace_view: command frontends.

def NN.Widgets.tapeHtml {α : Type} [TorchLean.Storage α] [ToString α] (t : Runtime.Autograd.Tape (Spec.SomeTensor α)) (maxDotChars : ℕ := 6000) :

ProofWidgets.Html

Render a tape as an HTML panel (nodes + DOT).

Instances For

source

def NN.Widgets.tapeGradsHtml {α : Type} [TorchLean.Storage α] [ToString α] [Add α] [One α] (t : Runtime.Autograd.Tape (Spec.SomeTensor α)) (outId : ℕ) :

ProofWidgets.Html

Render gradients from a scalar output id (like loss.backward()).

Instances For

Reverse-Pass Trace #

#tape_grads_view answers "which nodes received a gradient?".

When you need to understand why a gradient was produced (or why it is missing), it is useful to see the reverse traversal itself:

the upstream cotangent at each node, and
the per-parent contributions returned by the node’s local VJP rule.

This viewer runs reverse-mode, then recomputes each visited node's local VJP from its accumulated cotangent for display in reverse id order. It does not record intermediate accumulator updates.

source