Solving errors in AvgPool3d and Maxpool3d Implementations

[Shashi456]

the error inherently is :

error: cannot convert value of type '(Int, Int, Int, Int, Int)' to expected argument type '(Int, Int, Int, Int)'
        return input.maxPooled(kernelSize: poolSize, strides: strides, padding: padding)

I'm still working on how to solve it, Any suggestions are welcome :).

[Shashi456]

This should not have been merged because it does not compile. The self.init statement cannot appear on top level.

@rxwei solved this comment.

[Shashi456]

@jekbradbury @dan-zheng you mentioned in these comments about how maxPooled should probably be changed to a more abstract implementation. This compilation does throw an error because maxPooled is defined in (int, int, int, int) . Should i try to add an additional definition for maxpooled3d or is there any way to abstract the main maxpooled vjp

[dan-zheng]

Thank you for fixing the errors!

[dan-zheng]

@jekbradbury @dan-zheng you mentioned in these comments about how maxPooled should probably be changed to a more abstract implementation. This compilation does throw an error because maxPooled is defined in (int, int, int, int) . Should i try to add an additional definition for maxpooled3d or is there any way to abstract the main maxpooled vjp

Good catch! I wouldn't worry about those comments for now, actually.

Eventually, we should probably change APIs liked maxPooled and convolved2D to take 2 less stride/size arguments (e.g. so that convolved2D takes strides as (Int, Int) instead of (Int, Int, Int, Int)).

It would be good to implement that in a separate dedicated patch.

[Shashi456]

@dan-zheng, swift build still throws this error :

error: cannot convert value of type '(Int, Int, Int, Int, Int)' to expected argument type '(Int, Int, Int, Int)'
        return input.maxPooled(kernelSize: poolSize, strides: strides, padding: padding)

and I think we might have to create some sort of fix in this patch itself. I'm still wrapping my head around the error, but it makes sense since the vjp is written for (Int, Int, Int, Int) and not a higher dimension

[dan-zheng]

@dan-zheng, swift build still throws this error :

error: cannot convert value of type '(Int, Int, Int, Int, Int)' to expected argument type '(Int, Int, Int, Int)'
        return input.maxPooled(kernelSize: poolSize, strides: strides, padding: padding)

and I think we might have to create some sort of fix in this patch itself.

Aha, I see. The current maxPooled method is really maxPooled2D, and averagePooled is really averagePooled2D. As you observed, to implement AvgPool3D and MaxPool3D layers, it's necessary to first implement averagePooled3D and maxPooled3D methods.

averagePooled3D and maxPooled3D should be defined in Operators.swift. They should respectively call Raw.avgPool3D and Raw.maxPool3D, and their VJPs should respectively call Raw.avgPool3DGrad and Raw.avgPool3DGrad. (Raw operators are defined here.) The current averagePooled and maxPooled should be renamed to averagePooled2D and maxPooled2D.

Would you like to take this on? 🙂

[Shashi456]

@dan-zheng I will try, hope you help in debugging if i have doubts xD

[Shashi456]

@dan-zheng can you please review? sorry for the delay!

[dan-zheng]

The VJPs for averagePooled2D/3D and maxPooled2D/3D look good!

You also need to rename averagePooled and maxPooled to averagePooled2D and maxPooled2D and add averagePooled3D and maxPooled3D in the same file (Operators.swift). I don't believe the code currently compiles - could you please compile locally before committing code, to make sure it works?

Let us know if you have questions!

[Shashi456]

@dan-zheng i'm just getting this final error in the build, i don't know what to make of it.

[dan-zheng]

I'll take a look today Pawan. Thanks for your patience!

[Shashi456]

@dan-zheng is there any sort of rough roadmap for swift-apis i could look at?

[rxwei]

An ongoing plan is to finish #54, and the list of layers you provided there is great to follow!

API folks on the Swift for TensorFlow team (including @dan-zheng and myself) have been recently working on language/compiler enhancements (e.g. automatic differentiation). Once we finish the AD control flow work by the end of May, we'll be able to simplify this library by dropping most custom VJPs.

Meanwhile, @eaplatanios and @pschuh are driving #109, which is moving all tensor APIs from apple/swift into this repository. When that's finished, we will be able to iterate faster and improve the tensor library along with layers/optimizers that depend on them.

[Shashi456]

@rxwei thats amazing! With this PR and #75 we will be done with pooling layer implementations, from a usage point of view I guess the recurrent layers are the ones remaining. hope to keep contributing more :)

[dan-zheng]

LGTM! Thank you Pawan!