Seems bug in torch.mode

[lambdawill]

th> a = torch.rand(10):round()

th> a
 0
 1
 1
 1
 0
 1
 1
 0
 0
 1

th> y,i = torch.mode(a)

th> y
 1
[torch.DoubleTensor of size 1]

                                                                      [0.0001s]
th> i
 10
[torch.LongTensor of size 1]

When using 2D tensor

th> a = torch.rand(4,5):round()
                                                                      [0.0001s]
th> y,i = torch.mode(a)
                                                                      [0.0000s]

th> a
 0  1  0  1  1
 1  0  1  0  0
 0  0  0  1  0
 0  1  0  1  0
[torch.DoubleTensor of size 4x5]
                                                                      [0.0002s]

th> y
 1
 0
 0
 0
[torch.DoubleTensor of size 4x1]
                                                                      [0.0001s]

th> i
 5
 5
 5
 5
[torch.LongTensor of size 4x1]

[pavanky]

What is the bug here?

[lambdawill]

@pavanky I hope the answer to the first example should be

2
3
4
6
7
10

For the second example, all equals '5' ? It is quite strange! Maybe it should output

5
3
4
1

I cannot understand what the variable 'i' really means according to the output.

[pavanky]

@TYLVY It looks like it is returning the last instance of the mode inside the tensor. It so happens both of your examples have mode as the last element. May be try a few other examples?

[wickedfoo]

The choice of the index is completely arbitrary but correct. The guarantee it is providing is "some index whose value in the original tensor is equal to the mode", not the first index etc.