PyTorch-Traps

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18

>>> import torch as th
>>> zero = th.tensor([0])
>>> zero2d = th.tensor([0,0])
>>> empty = th.tensor([])
>>> if not zero:
...     print("zero is False")
... 
zero is False
>>> if not empty:
...     print("empty is False")
... 
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: Boolean value of Tensor with no values is ambiguous
>>> if not zero2d:
...     print("zero2d is False")
... 
RuntimeError: Boolean value of Tensor with more than one value is ambiguous

1
2

>>> a3d.shape, b2d.shape, c3d.shape, d2d.shape, e3d.shape
(torch.Size([2, 3, 4]), torch.Size([2, 3]), torch.Size([2, 1, 4]), torch.Size([3, 4]), torch.Size([2, 1, 3]))

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22

>>> (a3d+c3d).shape # [2, 3, 4]+[2, 1, 4]
torch.Size([2, 3, 4])
>>> th.eq(c3d.tile(1,3,1) + a3d, a3d + c3d).all().item()
True


>>> (a3d+b2d).shape # [2, 3, 4]+[2, 3]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: The size of tensor a (4) must match the size of tensor b (3) at non-singleton dimension 2


>>> (a3d+d2d).shape # [2, 3, 4]+[3, 4]
torch.Size([2, 3, 4])
>>> th.eq(a3d + d2d.unsqueeze(dim=0).tile(2,1,1), a3d + d2d).all().item()
True


>>> (e3d+b2d).shape # [2, 1, 3]+[2, 3]
torch.Size([2, 2, 3])
>>> th.eq(e3d.tile(1,2,1) + b2d.unsqueeze(dim=0).tile(2,1,1), e3d + b2d).all().item()
True