Remove original MultiVector class and use bare tf.Tensor instances and functions on them instead
Use dense representation instead of sparse with blade indices
Add new MultiVector class that just wraps those functions, operates on MultiVector instances and exposes less verbose operators (eg. ~ for reversion as before)
Advantages:
Dense is faster (STA MV*MV roughly 50x speedup on GPU)
MultiVector
class and use baretf.Tensor
instances and functions on them insteadMultiVector
class that just wraps those functions, operates onMultiVector
instances and exposes less verbose operators (eg.~
for reversion as before)Advantages:
MV*MV
roughly 50x speedup on GPU)tf.Tensor * MultiVector
issues (becausetf.Tensor
's__mul__
doesn't raiseNotImplemented
)Disadvantages:
tf.Tensor
representation and theMultiVector
wrapper