Read/Write tensors to HDF5 files

Question

The internal format used by "writeToFile" and "readFromFile" is not explicit. Is there any way of writing things e.g. using a HDF5 format? This would be useful for inspecting the results using hdfview or silx, and to interface ITensor with other MPS programs. Any development planned in this direction?

Answer 1

Yes, HDF5 would be great to have and we are planning to support HDF5 at some point soon, probably starting this summer. But I'm not sure when it will happen since I'm the only developer currently and I'll be traveling quite a bit until July.

In the meantime, if you want to create your own code to export ITensor data to a format of your choosing, here is a documentation page showing how to extract the storage inside of a dense ITensor, which is just a std::vector of Real (double) or Cplx (std::complex) numbers:
http://itensor.org/docs.cgi?page=formulas/extractdense

If you decide to work on that and encounter a more complicated ITensor storage type (such as the storage of an IQTensor) I can discuss with you how to extract that storage and what format it has.

Best regards,
Miles

Comments

Hi,
is there something analogous for IQTensor?

Best,
Mateusz

---

Hi Mateusz,
Yes you can extract the storage of any ITensor or IQTensor. Here is some code analogous to the code linked above but for extracting the storage of a dense, real IQTensor:

    auto extractReal = [](QDense<Real> const& d)
        {
        return d.store;
        };
    
    auto store = applyFunc(extractReal,T.store());

The variable "store" above will now hold a copy of the storage of the IQTensor "T". This storage is an object of type QDense<Real> (or alternatively QDenseReal which is an alias or typedef). I haven't gotten around to documenting the storage types in detail yet, but briefly the definition of QDense is in itensor/itdata/qdense.h and it has two main storage fields: one called "offsets" is a list of block information (the block number, when viewing an IQTensor as a "tensor of tensors" and the offset in the data storage array) and the storage array "store" which is just an array of numbers (real numbers in the case of QDenseReal).

Answer 2

So here's another piece of an answer: I wanted to write an ITensor of results to an hdf5 file for analysis in Julia. To save everybody else from having to re-invent this particular wheel, the code is in a gist. (Too long for the answer box.)

FWIW the code in the gist can take IQTensors---but it deals with them by flattening them down to ordinary ITensors.

Comments

(Looking over this again, I'm realizing the comments are a little unhelpful: e.g. it's useful to think about HDF5 doing reshaping, but I'm willing to bet an hdf5 dataset is just a big contiguous block of storage with metadata for stride etc.)