Computational graph with reverse-mode differentiation (Geron Ch 12, morie.fn hmcgrf)

Builds the expression DAG, evaluates it forwards, then makes ONE reverse adjoint sweep that yields every partial derivative at once – whatever the number of variables. A central-difference check at h = 1e-5 is returned alongside, so the analytic gradient is verified by an independent route rather than asserted.

Usage

morie_geron_computational_graph(expr, values = NULL)

Arguments

expr: Expression tree.
values: Named list of variable values.

Value

List with value, grad, gradient, node_grads, nodes, n_nodes, topo_order, fd_check, estimate, n.

Details

Expressions are nested lists: list("mul", "x", list("add", "y", 2)), with bare strings for variables and bare numbers for constants. Supported ops: add, sub, mul, div, pow, neg, exp, log, sin, cos, tanh, sqrt, square.

Examples

morie_geron_computational_graph(expr = 5L)
#> $value
#> [1] 5
#> 
#> $grad
#> list()
#> 
#> $gradient
#> list()
#> 
#> $node_grads
#> [1] 1
#> 
#> $nodes
#> $nodes[[1]]
#> $nodes[[1]]$op
#> [1] "const"
#> 
#> $nodes[[1]]$value
#> [1] 5
#> 
#> $nodes[[1]]$inputs
#> integer(0)
#> 
#> 
#> 
#> $n_nodes
#> [1] 1
#> 
#> $topo_order
#> [1] 0
#> 
#> $fd_check
#> list()
#> 
#> $estimate
#> [1] 5
#> 
#> $n
#> [1] 1
#> 
#> $method
#> [1] "expression DAG with forward evaluation and reverse-mode adjoint sweep"
#>