Reverse Rule

Construct a reverse-mode autodiff rule for a primitive. Provide exactly one of backward and forward.

Pass backward when the primitive's forward statement can run unmodified, which covers most use cases. It has the signature function(inputs, outputs, grads, params, required) and returns a list with one entry per input: that input's gradient, or NULL where required says it is not needed.

Pass forward when a slightly different forward pass enables a more efficient backward pass. It has the signature function(inputs, params, required), where required says which inputs need a gradient, and returns list(outputs = , backward = ): the forward results and a closure with the signature of backward above, which can use intermediate values of the forward pass via lexical scoping.

A rule with a forward pass can also be the rule of prim_custom_gradient().

Usage

rule_reverse(backward = NULL, forward = NULL)

Arguments

backward

(NULL | function)
Backward hook for the default case.

forward

(NULL | function)
Alternative forward hook that returns both the outputs and a backward closure.

Value

(anvl_rule_reverse)

See also

graph_gradient()

Examples

# the rule of prim_negate()
rule_reverse(function(inputs, outputs, grads, params, required) {
  list(if (required[[1L]]) prim_negate(grads[[1L]]))
})
#> $forward
#> NULL
#> 
#> $backward
#> function (inputs, outputs, grads, params, required) 
#> {
#>     list(if (required[[1L]]) prim_negate(grads[[1L]]))
#> }
#> <environment: 0x619e41eac0b0>
#> 
#> attr(,"class")
#> [1] "anvl_rule_reverse"