Pytaco API: Bad operand type for unary -: 'tensor'

[ByzanTine]

I am simply doing

>>> import pytaco as pt
>>> a = pt.tensor()
>>> -a
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: bad operand type for unary -: 'tensor'

Any hassle on supporting the unary operator?

[Infinoid]

What result would you expect?

The unary operator is defined on index expressions.

Example 1:

t = pytaco.tensor()
print(-t[None])

Output:

IndexExpr(-A0)

Example 2:

t = pytaco.tensor([10])
print(-t[5])

Output:

-0.0

[ByzanTine]

Thanks for the reply, I was expecting all values in the tensor to be negated, e.g like numpy API, where

>>> import numpy
>>> import numpy as np
>>> np.ones(5)
array([ 1.,  1.,  1.,  1.,  1.])
>>> -np.ones(5)
array([-1., -1., -1., -1., -1.])

[Infinoid]

You can do that with index notation. Taco is much more explicit about indexes than numpy is.

#!/usr/bin/env python3
from pytaco import tensor, index_var, from_array, Dense
import numpy

input = numpy.arange(20).reshape([5, 4])
x, y = [ index_var(c) for c in 'xy' ]

data = from_array(input)
negated = tensor([5, 4], [Dense, Dense])
print("data:", data)
negated[x, y] = -data[x, y]
print("negated:", negated)

Output:

data: A0 (5x4) ({dense},{dense}; 0,1):
dense (0): 
  [5]
dense (1): 
  [4]
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]
negated: A6 (5x4) ({dense},{dense}; 0,1):
dense (0): 
  [5]
dense (1): 
  [4]
[0, -1, -2, -3, -4, -5, -6, -7, -8, -9, -10, -11, -12, -13, -14, -15, -16, -17, -18, -19]

[ByzanTine]

Ofc, index notation works, but from a user perspective, dealing a negation directly with indices sounds cumbersome?
FWIW, currently, in our code, we were hacking like:

setattr(pt.tensor, '__neg__', lambda x: 0 - x)

[Infinoid]

Yeah, I see what you mean.

We've been talking about adding tensor-level operations, so that the syntax would look more like linear algebra... you could add or multiply vectors or matrices together without worrying about the indexes. Unary negation of a tensor would be something we add at the same time.

[rawnhenry]

Ah I don't think I overloaded negation in PyTaco. I can take a look at adding that since I think it should be quick.

Some operands in PyTaco emulate numpy behavior already by creating index notation for the user. I can add this for negation as well.

[rawnhenry]

This is addressed in #350.

You can now do:

#!/usr/bin/env python3
from pytaco import tensor, index_var, from_array, Dense
import numpy as np

arr_np = np.arange(20).reshape([5, 4])

data = from_array(arr_np)
print("data:", data)
print("negated:", -data)

To get the following output:

data: A0 (5x4) ({dense},{dense}; 0,1):
dense (0): 
  [5]
dense (1): 
  [4]
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]
negated: A6 (5x4) ({dense},{dense}; 0,1):
dense (0): 
  [5]
dense (1): 
  [4]
[0, -1, -2, -3, -4, -5, -6, -7, -8, -9, -10, -11, -12, -13, -14, -15, -16, -17, -18, -19]

[ByzanTine]

Thanks for the fix! Verified it works.