Usage example of RNN with large inputs about hype

[cgravill]

Perhaps we should add an option to control collecting history?

Hype/src/Hype/Optimize.fs

Lines 370 to 384 in 366daa7

	module Params =
	let Default = {Epochs = 100
	LearningRate = LearningRate.DefaultRMSProp
	Momentum = NoMomentum
	Loss = L2Loss
	Regularization = Regularization.DefaultL2Reg
	GradientClipping = NoClip
	Method = GD
	Batch = Full
	EarlyStopping = NoEarly
	ImprovementThreshold = D 0.995f
	Silent = false
	ReturnBest = true
	ValidationInterval = 10
	LoggingFunction = fun _ _ _ -> ()}

It could default on but allow people with large models to disable without having to recompile.

Does anyone have an optimization real use case they could share? We could potentially speed it up as well.

[zgrkpnr]

@adicirstei I can offer you my solution, but it requires you to modify source code.

In Optimize.fs find the line whist <- [w] @ whist and comment it out or delete it.
This line accumulates all the weight history which gets very large if your model is large. After that, it works great.

[adicirstei]

Thanks @zgrkpnr!

I'll give it a try. Hope I'll be able to compile it.

[smoothdeveloper]

@zgrkpnr do you remember what was the size of that list when you got OutOfMemory?

Looking at the code quickly, it feels this should actually be a ResizeArray (initialized with a size matching iters), this would also make the last step (where it is currently reversed and converted to array) faster.

This change could save roughly 4 bytes per element.

[zgrkpnr]

@smoothdeveloper Let's say our model has 900,000 parameters to optimize (which is quite normal for deep structures). 100 minibacthes and 10 Epochs for each brings us 100 x 10 x 900,000 single precision floating points of the size 3.4 b bytes. Roughly 3GB. I may be wrong, of course.

@adicirstei If the solution works, let me know.