torch.argmax() returning different indices to numpy.argmax() when the element values are the same

[leockl]

🐛 Bug

torch.argmax() returning different indices to numpy.argmax() when the element values are the same.

To Reproduce

Code:

import torch
import numpy as np

arr1 = np.array([1,2,3])
arr2 = np.array([5,10,3])
my_list_arr = [arr1, arr2]
A = np.transpose(my_list_arr)
A

Output:

array([[ 1,  5],
       [ 2, 10],
       [ 3,  3]])

Code:

tsr1 = torch.tensor(np.array([1,2,3]))
tsr2 = torch.tensor(np.array([5,10,3]))
my_list_tsr = [tsr1, tsr2]
B = torch.stack(my_list_tsr, dim = 1)
B

Output:

tensor([[ 1,  5],
        [ 2, 10],
        [ 3,  3]])

Code:

np.argmax(A, axis = 1)

Output:

array([1, 1, 0])

Code:

torch.argmax(B, dim = 1)

Output:

tensor([1, 1, 1])

Expected behavior

The returned indices should both be the same using numpy.argmax() or torch.argmax(), but the returned index for the 3rd element is different.

Environment

Collecting environment information...
PyTorch version: 1.5.1+cu101
Is debug build: No
CUDA used to build PyTorch: 10.1

OS: Ubuntu 18.04.3 LTS
GCC version: (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
CMake version: version 3.12.0

Python version: 3.6
Is CUDA available: Yes
CUDA runtime version: 10.1.243
GPU models and configuration: GPU 0: Tesla K80
Nvidia driver version: 418.67
cuDNN version: /usr/lib/x86_64-linux-gnu/libcudnn.so.7.6.5

Versions of relevant libraries:
[pip3] numpy==1.18.5
[pip3] torch==1.5.1+cu101
[pip3] torchsummary==1.5.1
[pip3] torchtext==0.3.1
[pip3] torchvision==0.6.1+cu101
[conda] Could not collect

Additional context

For consistency, would be helpful if torch.argmax() returns the same indices to numpy.argmax() when the element values are the same, where numpy.argmax() is the more commonly used function.

[kshitij12345]

Another Example:

>>> import torch
>>> x = torch.ones(3,3)
>>> x.argmin()
tensor(8)
>>> x.numpy().argmin()
0
>>> x.argmax()
tensor(8)
>>> x.numpy().argmax()
0

[colesbury]

Thanks for the report. This is the same underlying issue as #22853. Let's track this there.

[fdwr]

@leockl I wonder if you know any models that were affected by this tie-breaker behavior? I ask because we're deciding the default behavior of ties for WebNN's argMin/argMax (comparing numpy/PyTorch/TF/CoreML). TY.