Very weird behavior when converting to torch.tensor

1

I’m pretty new to PyTorch so, please excuse me if this question is too remedial. I have the following code for inference of my trained model which takes an image, does some pre-processing on it, converts to a tensor and finally performs a forward pass through the network.

   img = np.array(Image.open(img_path))

        orig_size = img.shape[:2][::-1]

        img_inp = torch.tensor(prepare_img(img).transpose(2, 0, 1)[None]).float()
        if has_cuda:
            img_inp = img_inp.cuda()

        segm = net(img_inp)[0].data.cpu().numpy().transpose(1, 2, 0)

I have successfully trained this segmentation model on my desktop machine. The inference code runs perfectly on it, but when I clone it to my laptop it gives me the following error:

       img_inp = torch.tensor(prepare_img(img).transpose(2, 0, 1)[None]).float()
RuntimeError: unsupported operation: more than one element of the written-to tensor refers to a single memory location. Please clone() the tensor before performing the operation.

I can bypass this error on my laptop by passing the result of prepare_img as a copy but that leads to memory leakages.

PS. On my Laptop I have built PyTorch from source for GPU compute and the torch version is 1.5.0a0+55c382e. I have a hunch that this is the problem. My desktop machine has PyTorch installed simply from anaconda’s repo.

Any help on this will be greatly appreciated. Thanks

2

Hi,

The problem is most likely that you try ot modify inplace a Tensor that does not support it. These checks were added fairly recently to prevent users from seeing behavior that are usually unexpected.

Could you give a small code sample that reproduces the error please?

3

Hey, @albanD thanks for the reply. You can check out the code in this repo. This is what I’m trying to re-purpose. Try running any one of the example notebooks. on PyTorch 1.5.0a0+55c382e

4

Seems like the main problem is when the torch.tensor is converted to float. When it’s removed at least the Run-time error is gone(appears instead when tensor is converted to .cuda()) but now how do I change to torch.float dtype?(already tried .type(torch.float))

5

I think .float only works for scalarr, it’ll return the value as a python float. If u want to convert types for tensors, try a = a.to(dtype=tc.flaot)

6

@G.M Thanks for the reply. Still the same problem:

  img_inp = torch.tensor(prepare_img(img).transpose(2, 0, 1)[None]).to(dtype=torch.float)
RuntimeError: unsupported operation: more than one element of the written-to tensor refers to a single memory location. Please clone() the tensor before performing the operation.
7

How come this works?

import numpy as np
a = np.ones(5)
b = torch.from_numpy(a).float()   # because we change the dtype 'b' contains a copy of  'a' 
np.add(a, 1, out=a) 

print(a, a.dtype)
>>> [2. 2. 2. 2. 2.] float64

print(b, b.dtype)
>>> tensor([1., 1., 1., 1., 1.]) torch.float32

Secondly, this works as expected:

import numpy as np
a = np.ones(5)
b = torch.from_numpy(a)
np.add(a, 1, out=a) # changed to np array, changes the torch.tensor

print(a, a.dtype)
>>> [2. 2. 2. 2. 2.] float64

print(b, )
>>> tensor([2., 2., 2., 2., 2.], dtype=torch.float64)
8

Do u have any in-place operations or operations like expand in ur code? I’ve experienced the issue with the usage of expand.

9

Nope. Can’t find any expand commands. You can check out the repo here

10

Can u try creating the tensor like this?

img_inp = torch.from_numpy(prepare_img(img).transpose(2, 0, 1)[None]).to(dtype=torch.float,device=UR_DEVICE)
11

Hi,

The issue comes from the [None] I think. You change this line to first convert to a Tensor then .unsqueeze(0) to get the exact same result:

img_inp = torch.tensor(prepare_img(img).transpose(2, 0, 1)).float().unsqueeze(0)
1 Like
12

Thanks a bung @albanD got it working