ad-dual - [no description]

	Commit message (Collapse)	Author	Age
*	Dual arrays is >100x faster than 'ad' on large fneuralHEAD master	Tom Smeding	24 hours
\| \| \| \| \| \| \| \| \| \| \| \| \|	ad:Numeric.AD.Double / ad-dual:Numeric.ADDual.Array.Internal Prelude> 1.129e-3 / 41.89e-6 -- neural-100 26.951539746956314 Prelude> 34.67e-3 / 156.9e-6 -- neural-180 220.9687699171447 Prelude> 79.03e-3 / 178.6e-6 -- neural-500 442.4972004479283 Prelude> 365.3e-3 / 665.5e-6 -- neural-2000 548.9105935386928
*	WIP array stuff untested	Tom Smeding	33 hours
\|
*	WIP array AD	Tom Smeding	35 hours
\|
*	Remove redundant Num constraint on writeTape	Tom Smeding	2 days
\|
*	Optimise by backpropagating scalar computation in C	Tom Smeding	2 days
\|
*	Working tests and benchmarks against 'ad'	Tom Smeding	4 days
\|
*	Neural returns 1 on all-zero input, not 0.	Tom Smeding	4 days
\|
*	Evaluate result Dual before backpropagating	Tom Smeding	4 days
\| \| \| \| \|	This ensures that the tape is actually created/written before we start backpropagating over it, lol.
*	Nicer API for writing Dual instances	Tom Smeding	4 days
\|
*	Add (toggleable) debug code	Tom Smeding	4 days
\|
*	Try debugging crash	Tom Smeding	5 days
\|
*	Attempt at a benchmark (crashes)	Tom Smeding	6 days
\|
*	Initial	Tom Smeding	6 days