torch.addmm接收到无效的参数组合

Question

In the official webpage of pytorch I saw the following code with answers:

>> a = torch.randn(4, 4)
>> a

0.0692  0.3142  1.2513 -0.5428
0.9288  0.8552 -0.2073  0.6409
1.0695 -0.0101 -2.4507 -1.2230
0.7426 -0.7666  0.4862 -0.6628
torch.FloatTensor of size 4x4]

>>> torch.max(a, 1)
(
 1.2513
 0.9288
 1.0695
 0.7426
[torch.FloatTensor of size 4]
,
 2
 0
 0
 0
[torch.LongTensor of size 4]
)

I know that the first result corresponds to the maximum number per row, however I do not get the second tensor (LongTensor)

I tried other random example and after a pytorch.max, I came to find these results

0.9477  1.0090  0.8348 -1.3513
-0.4861  1.2581  0.3972  1.5751
-1.2277 -0.6201 -1.0553  0.6069
 0.1688  0.1373  0.6544 -0.7784
[torch.FloatTensor of size 4x4]

(
 1.0090
 1.5751
 0.6069
 0.6544
[torch.FloatTensor of size 4]
, 
 1
 3
 3
 2
[torch.LongTensor of size 4]
)

Can anyone tell me what does it really mean these LongTensor data? I thought it was a strange casting between tensors, however after a simple casting of a float tensor, I see that it just cuts decimals

Thanks

Answer 1

It just tells the index of the max element in your original tensor along the queried dimension.

Eg

0.9477  1.0090  0.8348 -1.3513
-0.4861  1.2581  0.3972  1.5751
-1.2277 -0.6201 -1.0553  0.6069
 0.1688  0.1373  0.6544 -0.7784
[torch.FloatTensor of size 4x4]

# torch.max(a, 1)
(
 1.0090
 1.5751
 0.6069
 0.6544
[torch.FloatTensor of size 4]
, 
 1
 3
 3
 2
[torch.LongTensor of size 4]
)

In the above example in torch.LongTensor ,

1 is the index of 1.0090 in your original tensor (torch.FloatTensor)
3 is the index of 1.5751 in your original tensor (torch.FloatTensor)
3 is the index of 0.6069 in your original tensor (torch.FloatTensor)
2 is the index of 0.6544 in your original tensor (torch.FloatTensor)

along dimension 1 .

---

Instead, if you'd have requested torch.max(a, 0) , the entries in torch.LongTensor would correspond to the indices of max elements in your original tensor along dimension 0 .