search

jcmgray / quimb

A python library for quantum information and many-body calculations including tensor networks.
http://quimb.readthedocs.io
Other
501 stars 109 forks source link

tensor_contract() interface prevents turning off contraction path cache #112

Open mattorourke17 opened 2 years ago

mattorourke17 commented 2 years ago

When calling TensorNetwork.contract() or any related function that filters down to tensor_contract(), I don't think it is currently possible to shut off caching of contraction expressions. If the user supplies cache=False as a kwarg that eventually reaches tensor_contract(..., **contract_opts), it gets passed to get_contraction(eq, *shapes, cache=True, get='expr', optimize=None, **kwargs) in the kwargs, whereas cache is already manually set in the function call. The body of get_contraction() does not catch the case that a cache value is passed in **kwargs, and ends up always passing True as the cache value when calling _get_contraction().

It is important to be able to shut off caching from the high-level interfaces because sometimes a user might want to contract two networks with the same opt_einsum expression but different values of the bond dimensions, in which case opt_einsum throws an error like ValueError: Size of label 'g' for operand 6 (2) does not match previous terms (3)

I think this can either be fixed by checking if 'cache' in kwargs in get_contraction() or by checking if 'cache' in contract_opts in tensor_contract()

mattorourke17 commented 2 years ago

Looking closer at opt_einsum, I actually don't think the error I quoted is related to the cache kwarg propagation issue

jcmgray commented 2 years ago

quimb should cache on both the contraction equation and sizes - and the returned expression should be able to handle any dimensions changing - so not totally clear the source of the error, but it implies that two different tensors in the same contraction have index dimensions that don't match.

Having said that, I have some refactoring of the contraction parsing stuff that I need to push, including optionally using cotengra rather than opt_einsum, to perform the contraction which has various advantages and might be easier to understand.