malmaud
Implement eigsum
currently TensorFlow.jl is incompatible with TensorOperations.jl, have you considered enacting (multiple-legs) tensor contraction in TensorFlow.jl, as the counterpart of tf.tensordot in Python API? thank you a lot!
I'd welcome a PR for this. Probably not something I'd get around to myself for a while, though :(
If you need it, may I suggest helping out by implementing it and contributing it back to the community?
It is a kinda odd operation for TensorFlow. because TensorFlow is not actually really about Tensors, it is about Flow (as an odd name for computational graphs), where the operations happen the be on Arrays of n-dimensions (i.e. Tensors) -- but n is normally 2. But these operations are not really Tensor Operations as such. Bastically TensorFlow is not a Tensor Toolkit. It is just badly named.
vs TensorOperations.jl which is actually a Tensor math toolkit. As such I feel like compatibility with TensorOperations is not exactly a priority feature. They are different libraries with very different target userbase, and thus different priority features.
As hinted at in the OP (and for my own future reference) eigsum and tensordot is implemented in the python client. Not in the TensorFlow backend.
It is certainly implementable in TensorFlow.jl It is a combination of reshape, concat, MatMul and maybe permute_dims.
