TensorSafe

Encode tensor/matrix shapes into types

Motivation

Writing tensor/matrix code is error-prone. Because the compiler doesn't stop you from writing shape/dimension-related errors like multiplying two matrices with incompatible shapes, or forgetting transpose a matrix before an addition, etc. Instead, it lets your program run and run... oops! ... suddenly crash. I can recall myself wasting quite a lot of time debugging problems like this. Wouldn't it be nice if we can detect and eliminate such errors at compiling time?

That's why you need TensorSafe, a tensor manipulation wrapper library for nd4j. TensorSafe encodes tensor shape information into their types and uses type-level programming (via Scala's implicit parameter resolution mechanism) to detect inappropriate computations at compiling time, moreover, it also computes the resulting shapes for you.

Basic Example

Define the Dimensions

First, let's define the Dimensions we are going to need in this example.

A dimension is a type that extends the Dimension trait, and a user defined dimension should extends VarDim, a subtrait of Dimension which means "Variable Dimension".

packagetensorsafe.exampleimporttensorsafe._traitDataNumberextendsVarDimtraitDataDimensionextendsVarDimtraitFeatureDimensionextendsVarDimobjectBasicExample {
defmain(args: Array[String]):Unit= {
// We'll write our tensors here
}
}

Define the tensors

With these dimensions in our hands, now we can define some tensors.

defmain(args: Array[String]):Unit= {
valinputData=TensorBuilder[RNil~DataNumber~DataDimension].ones
valweights=TensorBuilder[RNil~DataDimension~FeatureDimension].rand
valfeatureVectors= inputData * weights // '*' means matrix multiplication
}

Here we use TensorBuilder to create our tensors. The strange-looking type parameter RNil~DA~DB simply means a tensor of shape DA * DB. RNil marks the start of a type list, just like the Nil object always ends a normal scala List. But this type list grows rightward, that's why it's called a RList, and hence this so-called RNil.

By the way, ~ is just a type alias for tuple2, so A~B is equivalent to (A,B).

What's the tensor shape? Ask your IDE

Now let's check the type of these tensors! I can easily find them out by asking my IDE.

(In Intellij IDEA, it's View|TypeInfo)

You see? featureVectors has the type Tensor[((RNil,DataNum),FeatureDimension)], which means it's a tensor of shape DataNum * FeatureDimension, as expected.

So in TensorSafe, you don't need to manually remember and track the shapes of every tensor any more!

Provide DimValues

But wait! How dose the compiler know what values those dimensions should have?

If you run the code now... You'll see the following compile-time error:

Error:(14, 34) Not enough value information for Shape 'tensorsafe.~[tensorsafe.~[tensorsafe.RNil,tensorsafe.example.DataNumber],tensorsafe.example.DataDimension]'
Please make sure that each dimension has a corresponding DimValue in scope. Zero-rank Tensor is not allowed.
val inputData = TensorBuilder[RNil~DataNumber~DataDimension].zeros

Which basically says you must provide an implicit DimValue for each dimension you use in the TensorBuilder.

Let's fix this.

defmain(args: Array[String]):Unit= {
importDimValue.constimplicitvaldataNumber= const[DataNumber](10)
implicitvaldataDimension= const[DataDimension](3)
implicitvalfeatureDimension= const[FeatureDimension](6)
valinputData=TensorBuilder[RNil~DataNumber~DataDimension].ones
valweights=TensorBuilder[RNil~DataDimension~FeatureDimension].rand
valfeatureVectors= inputData * weights
println(featureVectors)
}

Now the program compiles, and print out the result.

Broadcasting

TensorSafe also supports numpy-like broadcasting.

In the above example, we can add a 1D bias vector to the 2D featureVectors:

valbias=TensorBuilder[RNil~FeatureDimension].randGaussian
valfeatureVectors= inputData * weights + bias

It also works for cases with higher ranks:

valt1=TensorBuilder[RNil~DataNumber~UnitDim~FeatureDimension].zeros // UnitDim is a dimension with value 1valt2=TensorBuilder[RNil~DataDimension~UnitDim].ones
valt3= t1 + t2 // t3 has type Tensor[RNil~DataNumber~DataDimension~FeatureDimension]

A more complex example

You can take a look at a neural network example from the file NNExample.scala