7
votes

Seems simple enough. I would have thought some kind of casting would be possible, but I can't seem to find any documentation for it.

While I have found ways in my application to avoid using the Eigen::Matrix class, TensorFlow only works with Eigen::Tensor, and another library I use only has functionality for working directly with Eigen::Matrix. It would be spectacular for code readability if I could cast a Tensor as a Matrix and work with that.

edit: it seems that TensorFlow DOES have a function to get an Eigen::Matrix out (still testing it out). Maybe that makes the original question less interesting (maybe no one NEEDS to convert Tensors to Matrices.) However I still think it's a valid question to ask. so I won't put down my

edit 2: going through the TF documentation after some build errors, it seems that tensorflow's Tensor::matrix() function simply returns a 2d Eigen::Tensor, so the conversion is in fact necessary.

1

1 Answers

13
votes

This is a common use case for TensorFlow's linear algebra ops, and an implementation can be found in tensorflow/core/kernels/linalg_ops_common.cc. However, that code is highly templatized, so it might be useful to have a concrete example.

Assuming you start with a tensorflow::Tensor called t with element type float, you can make an Eigen matrix m as follows:

tensorflow::Tensor t = ...;

auto m = Eigen::Map<Eigen::Matrix<
             float,           /* scalar element type */
             Eigen::Dynamic,  /* num_rows is a run-time value */
             Eigen::Dynamic,  /* num_cols is a run-time value */
             Eigen::RowMajor  /* tensorflow::Tensor is always row-major */>>(
                 t.flat<float>().data(),  /* ptr to data */
                 t.dim_size(0),           /* num_rows */
                 t.dim_size(1)            /* num_cols */);

If your tensor comes from the input of a tensorflow::OpKernel (e.g. in the Compute() method), you would use a slightly different type with the appropriate const qualification:

OpKernelContext* ctx = ...;
const tensorflow::Tensor t = ctx->input(...);

const auto m = Eigen::Map<const Eigen::Matrix<
                   float,           /* scalar element type */
                   Eigen::Dynamic,  /* num_rows is a run-time value */
                   Eigen::Dynamic,  /* num_cols is a run-time value */
                   Eigen::RowMajor  /* tensorflow::Tensor is always row-major */>>(
                       t.flat<float>().data(),  /* ptr to data */
                       t.dim_size(0),           /* num_rows */
                       t.dim_size(1)            /* num_cols */);