PyTorch - multiplying tensor with scalar results in zero vector

Question

I have no idea why the result is all 0 with tensor. Anything wrong here?

>>> import torch
>>> import numpy as np
>>> import math

>>> torch.__version__
'0.4.1'
>>> np.__version__
'1.15.4'

>>> torch.arange(0, 10, 2) *-(math.log(10000.0) / 10)
tensor([0, 0, 0, 0, 0])
>>> np.arange(0, 10, 2) *-(math.log(10000.0) / 10)
array([-0.        , -1.84206807, -3.68413615, -5.52620422, -7.3682723 ])

>>> torch.arange(0, 10, 2)
tensor([0, 2, 4, 6, 8])
>>> np.arange(0, 10, 2)
array([0, 2, 4, 6, 8])

Answer 1

As written in the comment when using 0.4.0 get the same results as with numpy:

tensor([-0.0000, -1.8421, -3.6841, -5.5262, -7.3683])

However with 0.4.1 I'm getting a zero vector too.

The reason for this is that torch.arange(0, 10, 2) returns a tensor of type float for 0.4.0 while it returns a tensor of type long for 0.4.1.

So casting your tensor to float should work for you:

torch.arange(0, 10, 2).float() *-(math.log(10000.0) / 10)

Multiplying long and float works by heavy rounding, as the result is still a tensor of type long. So when converting a FloatTensor to a LongTensor values between -1 and 1 will be rounded to 0.

Since -(math.log(10000.0) / 10) results in -0.9210340371976183 your result is 0. So effectively -0.9210340371976183 is converted to type long before multiplying. But when converting it will be round down to 0, see this example:

t = torch.tensor((-(math.log(10000.0) / 10)))
print('FloatTensor:', t)
print('Converted to Long:', t.long())

Outout:

FloatTensor: tensor(-0.9210)
Converted to Long: tensor(0)

Thus:

torch.arange(0, 10, 2).float() *-(math.log(10000.0) / 10)

becomes:

torch.arange(0, 10, 2).float() * 0

Therefore you get a tensor of zeros as result.

---

Some more examples:

If you multiply it with a value between 1 and 2, lets say 1.7, it will always been rounded down to 1:

t = torch.tensor(range(5), dtype=torch.long)
print(t)
print(t * 1.7)

Output:

tensor([ 0,  1,  2,  3,  4])
tensor([ 0,  1,  2,  3,  4])

And similarly when multiplying with 2.7 results in an effective multiplication of 2:

t = torch.tensor(range(5), dtype=torch.long)
print(t)
print(t * 2.7)

Output:

tensor([ 0,  1,  2,  3,  4])
tensor([ 0,  2,  4,  6,  8])