torch.nn.MultiLabelMarginLoss has different performance on CPU and GPU.

[triumph-wangyuyang]

🐛 Describe the bug

Test on the CPU:

import torch
loss = torch.nn.MultiLabelMarginLoss()
x = torch.randn((1, 4))
y = torch.randint(4, 16, [1, 4])
out = loss(x, y)
print(out)

result: RuntimeError: argument #2 'target' is out of range

Test on the GPU:

import torch
loss = torch.nn.MultiLabelMarginLoss().cuda()
x = torch.randn((1, 4)).cuda()
y = torch.randint(4, 16, [1, 4]).cuda()
out = loss(x, y)
print(out)

result: tensor(5.7001, device='cuda:0')

torch.nn.MultiLabelMarginLoss has different performance on CPU and GPU.

Versions

pytorch: 1.12.1
Python version: 3.8
CUDA/cuDNN version: cuDNN 11.3
GPU models and configuration: RTX3060
Operating System：Windows

[triumph-wangyuyang]

Is there any progress on this issue?

[akashdhamasia12]

According to the definition as given here:
https://pytorch.org/docs/stable/generated/torch.nn.MultiLabelMarginLoss.html

y has to be in the range as:

thats why you are getting the error.

[triumph-wangyuyang]

According to the definition as given here: https://pytorch.org/docs/stable/generated/torch.nn.MultiLabelMarginLoss.html

y has to be in the range as:

thats why you are getting the error.

I know that parameters have this limitation, but I tested with unreasonable parameters to test whether the operator has a check when inputting unreasonable parameters. On this operator, I found that when the operator is running, and It is unreasonable to let the program continue without first checking whether the parameters are reasonable. Therefore, I suggest adding a check on the parameter size at the program entry, so that when it is unreasonable, an exception is directly thrown or a warning message is given.