flare.optimize
ada-delta
(ada-delta eta gamma epsilon)(ada-delta factory eta gamma epsilon)
batch-norm-clip-grad
(batch-norm-clip-grad model batch-size grad-clip)
DiffFn
protocol
members
val-at
(val-at this xs)
return [f(x) grad-f(x)] pair as double/double-array
init-model!
(init-model! optimizer model)
loss-fn
(loss-fn model build-graph data)
Optimizer
protocol
members
init
(init this params-node)
return state required for future computations
update-params!
(update-params! this params-node state)
write to :value of params node and return updated state
rand-bump-test
(rand-bump-test diff-fn xs)
reset-batch!
(reset-batch! optimizer model)
Reset all gradients except the target node
update-expected-sqs
(update-expected-sqs accum gamma update)
update-model!
(update-model! optimizer model)