It would be great to support basic cross-type auto-converting arithmetic operations for floating point types as well.

Copy/Assignment operations are already supported (VERY permissive):

f = torch.FloatTensor(5)
d = torch.DoubleTensor(5).fill_(float('nan')) # fill with nan to challenge double->int...
i = torch.IntTensor(5)
d.copy_(f)
f.copy_(d)    # ok, maybe I accept this as explicit cast...
f[:] = d[:]   # implicit conversion to smaller type? data-scientists world...
d[:] = f[:]
i[:] = d[:]   # wow, even double to int without complaints...

Things become strange/normal (depending on whether you are data or computer scientist) when you try to assign a SINGLE float to an int:

>>> i[1] = d[1]
RuntimeError: can't assign a float to a scalar value of type int

A nice feature I would like to see in the future are relatively save cross-type arithmetic ops for +,-,*,/ ... e.g.

r = d[:] + f[:]
# currently only possible with explicit conversion (-> additional copy), e.g.
r = d[:] + f[:].double()

Currently yields:

    return self.add(other)
TypeError: add received an invalid combination of arguments - got (torch.FloatTensor), but expected one of:
 * (float value)
      didn't match because some of the arguments have invalid types: (torch.FloatTensor)
 * (torch.DoubleTensor other)
      didn't match because some of the arguments have invalid types: (torch.FloatTensor)
 * (float value, torch.DoubleTensor other)
 * (torch.DoubleTensor mat1, torch.SparseDoubleTensor mat2)
 * (torch.DoubleTensor mat1, float value, torch.SparseDoubleTensor mat2)

Support multi-type operations where reasonable #28

Description

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions