Re: Looking for: tensor-library for sparse tensor based triple store

RESCAL? https://github.com/mnick/rescal.py

Best,
Jörn

> On 18 May 2017, at 18:28, Alexander Bigerl <bigerl@informatik.uni-leipzig.de> wrote:
> 
> Hi everyone,
> 
> I am working on a tensor-based triple store to query triple patterns (not full SPARQL). Therefor I'm looking for a suitable library supporting sparse tensor product. The programming language doesn't matter. But it would be nice if it was optimized for orthonormal-based tensors (means it doesn't need to distinguish between co- and contravariant dimensions for multiplication).
> 
> In more detail:
> 
> I represent my my data like this:
> 
>  • I have tensors storing boolean values. 
> 
>  • They are n >= 3 dimensional and every dimension has the same size m>1000000. 
> 
>  • Every dimension uses a natural number index 0...m.
> 
>  • The tensors are orthonormal-based so I don't need to distinguish between co- and contraviarant dimensions. 
> 
>  • There are only very few true values in every tensor, so the rest of the values is false. Therefor it should be sparse. Non-sparse is no option because of at least 1000000^3 entries. 
> 
> I'm looking for: 
> 
>  • efficient sparse n-D tensor implementation with support of a fast inner product like: Tαγβ • Dβδε = Rαγδε
> 
>  • optional: support for pipelining multiple operations
> 
>  • optional: support for logical and or pointwise multiplication of equal-dimensioned tensors.
> The following libraries don't do the trick for reasons:
> 
>  • Tensor flow: misses multiplication with non-dense-none-2D-matrices
>  • scipy sparse: supports only 2D representation and would output a dense narray for dotproduct
>  • theano: supports only 2D sparse tensors
>  • Shared Scientific Toolbox and Universal Java Matrix Package: don't support multiplication of n-D sparse tensors
> Who is wandering now where the triples are: They are mapped to the dimensions' index so that the coordinates of a true in a 3D Tensor represents a triple.
> 
> I would be very thankful for any comments or recommendations. 
> 
> Kind regards,
> 
> Alexander Bigerl
>

Received on Friday, 19 May 2017 14:10:47 UTC