RuntimeError: Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor for argument #2 'weight', help me understand how and when to use cuda

1

Hi, I got the above error, and from the docs I’m not able to understand where should I adjust to cuda.
For example in the docs they are showing how to create a tensor on the gpu, but what with the case when I’m loading Images? still the same?

Now I have a situation my tensors are cuda type, but my layers not, and that’s why I’m getting :
"Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor ".
Now I need to put .cuda() on each layer?(or something like this) doesn’t make sense.

This is were the program stoped, in the nn.conv2d , tried to “.cuda()”
, didn’t help

             return nn.Sequential(
                nn.Conv2d(inp, oup, 3, stride, 1, bias=False).cuda(),
                nn.BatchNorm2d(oup),
                nn.ReLU(inplace=True)
            )

I tried few of the solutions online, but nothing worked.

2

Your data as well as your model should be both on the GPU.
To transfer both you can use something like this:

model = model.to('cuda')
data = data.to('cuda')
target = target.to('cuda')

You don’t need to push each layer onto the GPU. The model will push all it’s layers to the device.

3 Likes
3

I have already do this job, but it doesn’t work.It seems the first batch can work well while the next gets error

Snipaste_2018-11-03_19-24-47
Snipaste_2018-11-03_19-24-471197×599 31.6 KB

4

Could you post your code so that we could have a look?
The issue is strange, as the first batch seems to work.
If you can’t post code due to some restrictions, could you come up with a small dummy code snippet to reproduce this issue?

5

I found the bug when I was copying my code.this is my test data that not transfered lead to the error.
Anyway, thanks for your attention.:blush:

1 Like
6

could you please tell how to solve this problem?